In the field of computer science, distributed systems are a challenging research area and an essential optimization practice for Internet applications, while CAP theory and BASE theory are two key concepts in distributed systems.

2. What is a distributed system

First, let’s talk about distributed systems. You can think of a distributed system as a large network of computers, consisting of multiple computer or server nodes, which may be located in different geographical locations.

As shown in the figure, the three nodes of the application layer are published in different cities. These nodes can communicate and collaborate with each other to accomplish complex tasks.

Imagine you are a team leader with a task to accomplish. If you complete it alone, it may take a long time.

But if you break the task down into several subtasks and assign them to your team members, they can work in parallel and complete the task faster. This is the core idea of distributed systems.

3 CAP Theory

Next, let’s talk about CAP theory, which is a very important principle in the design of distributed systems.

CAP refers to the three basic principles of Consistency, Availability, and Partition tolerance in distributed systems.

C - Consistency

Consistency means that no matter which node of the distributed system you read from, you get the same copy of the data and it ensures the accuracy of the data.

In distributed systems, there are three broad types of consistency which are Strong Consistency, Weak Consistency and Final Consistency.

Strong Consistency

Strong consistency requires that when a user accesses data in a distributed system, the data should be exactly the same regardless of the response from any node.

For example, there are 10 pairs of sneakers left in the inventory in the order system, Zhang San just bought a pair of sneakers, after the data update is completed, the next number of sneakers seen by Li Si is only 9 pairs, otherwise there may be overselling.

However, this requires more time and effort to coordinate, just like when Li Si is buying shoes, he must queue up and wait for Zhang San’s purchase action to finish first before he can continue, which is less efficient.

Weak Consistency

Weak consistency means that after the data in a distributed system has been updated, it is also allowed to allow subsequent accesses to get the old data before the update.

It is like going to a party where everyone has their own clock. The time of each clock may be a little different, but that doesn’t stop everyone from getting together and having fun.

Weak consistency improves the efficiency of the business, but sometimes it can lead to some confusion; imagine the long wait if the party-goers’ clocks are too far off.

Final Consistency

Final consistency is a special form of weak consistency that requires that the system’s data update is complete, and after a period of time, all subsequent accesses get the latest data.

It’s like message propagation in the circle of friends. When you send a message, it is not immediately seen by all your friends, but eventually, everyone will see the same message.

The vast majority of general business systems use final consistency as the design philosophy for distributed systems based on cost-effective considerations.

Consistency in CAP theory, on the other hand, requires strong consistency. As described in the official documentation: All nodes see the same data at the same time.

A - Availability

Availability means that every request of a distributed system should be responded to and should be completed within a limited time.

Availability ensures the stability and reliability of the system, and it describes the ability of the system to serve the user well, without user actions failing or access timing out, which affects the user experience.

It is the official term Reads and writes always succeed, the service is always available within the normal response time.

P - Partition Tolerance

Partition Tolerance means that the system can continue to run in case of network partition or communication failure, that is, if the network communication between nodes fails, or one of the nodes in the system has a problem, we still need to ensure that the business system is available.

That is, The system continues to operate despite arbitrary message loss or failure of part of the system, the distributed system is still able to externally provide services that satisfy consistency or availability in the event of a node or network partition failure.

4. Characteristics of CAP

4.1 Importance of partition fault tolerance

At this point, students with a basic knowledge of distributed computing may ask, “CAP theory is indeed very important, but it seems that these three characteristics cannot be satisfied at the same time,” right?

Yes, this is the core idea of CAP theory.

CAP theory tells us that in a distributed system, we can only satisfy at most two of the characteristics at the same time, but not all three at the same time.

Why can’t C, A, and P be both? First of all, we have to know that in a distributed system, due to the unreliability of the network, in order to ensure that the service can always be provided to the outside world, so ** partition fault tolerance must be guaranteed **.

Imagine if there is only one partition, talk about distributed is meaningless. And more than one partition, there must be partition failure problems, distributed systems to ensure partition fault tolerance has become the most basic claim.

So now we only need to consider whether we can satisfy both consistency and availability on the basis of partition fault tolerance, which we can prove by using the counterfactual method.

4.2 AP Or CP

Suppose we now have two partitions, P1 and P2, with the same data, D1 and D2, on the partitions, which are now identical.

Next, a request 1 accesses P1 and changes the data on D1. Then another request 2 accesses P2 to access the same data on D2.

At this point, we need to make a tradeoff.

Ensure consistency first

If we first guarantee to satisfy consistency and partition fault tolerance, i.e. CP.

This process can easily occur: D1 has updated the data, but when querying D2, the data is returned as old.

In order to ensure that D2 and D1 data are completely consistent, you must lock the D2 data on P2 when updating the D1 data, and then synchronize the D2 update after the D1 update is complete.

In this process, the locked D2 will not be able to give a real-time response to request 2, that is, violating the availability of P2.

Therefore, under the premise of consistency, CAP can not be satisfied at the same time.

Guarantee availability first

If availability and partition fault tolerance are guaranteed to be met first, that is, AP.

Availability requires that both P1 and P2 can respond in real time, so when D2 has just finished updating and has not yet been synchronized to D1, the data in the two DBs is inconsistent, which violates the data consistency on P1 and P2.

Therefore, under the premise of availability, CAP cannot be satisfied at the same time.

4.3 CAP tradeoffs

CAP You can’t have all three, so what should you choose? Generally according to our business we can have the following choices.

Satisfy consistency and partition fault tolerance CP

Guarantees strong consistency of partitioning (C) and does not require availability (A).

It is equivalent to waiting for the data to be fully synchronized before the request arrives at a certain system before getting the data response from the system. Generally, this model is used in financial systems where the data needs to be strictly consistent.

Meet Availability and Partition Fault Tolerance AP

Availability of partitions is guaranteed (A) and strong consistency is not required (C).

When requesting access to data in a partition, you may get old data that is not synchronized. This model generally only requires that the data meets the final consistency, which in turn ensures system response speed and high availability.

AP is widely used in the industry, such as the famous BASE theory (discussed in detail below).

Satisfying Availability and Consistency AC

As mentioned above, it is not possible to guarantee both strong consistency (C) and availability (A) of a system in a distributed system.

This is because partitioning in a distributed system is objectively unavoidable, whereas databases in a monolithic system can guarantee data consistency and availability through transactions, such as the four main characteristics of transactions (Atomicity, Consistency, Isolation, and Durability, or ACID for short) in MySQL.

5. BASE Theory

The BASE theory summarizes the practice of distributed systems on the Internet today. Its core idea is that since the cost of achieving strong consistency in a distributed system is too high, it is better to settle for the second best.

It is only necessary for each application partition to ensure data consistency to the best of its ability on the basis of providing highly available services, that is, to ensure the final consistency of the data.

BASE theory is the CAP to ensure partition fault tolerance (P) under the premise of availability (A) and consistency (C) of the trade-offs, which consists of Basically Available (Basic Available), Soft State (Soft State), Eventually-Consistent (Final Consistency) three aspects of the composition, referred to as the BASE Theory.

In distributed systems, CAP theory provides a theoretical framework, while BASE theory provides a guiding principle for practical operation.

5.1 Basic Availability

BASE theory recognizes that a distributed system may choose to reduce performance or consistency requirements in order to maintain basic availability in the face of failures or anomalies.

This means that the system may experience some transient inconsistencies, but will eventually reach a consistent state.

Just as a system design for a banking system generally has functional and non-functional requirements, we first need to ensure basic usability of the core functional requirements.

Functional Requirements

In a banking system, transaction modules such as user withdrawals, transfers, etc. are the core functions, which are the basic needs of users and cannot go wrong.

The non-core functions can be abnormal, but need to be guaranteed to be fixed within a period of time.

Non-functional requirements

Non-functional requirements refer to other requirements that user’s business does not depend on, such as performance-related: users are required to transfer money within 0.5 seconds, but due to network delays and other reasons, the response can be delayed to 1~2 seconds.

Due to such anomalies in the system, thus affecting the high availability of the system, but the core process is still available, i.e. basic availability.

5.2 Soft State

Soft state means that system services may be in an intermediate state, and data may be delayed in synchronization in the process of ensuring consistency, but it will not affect the availability of the system.

For example, after we have finished paying for a train ticket, we may be in an intermediate waiting state that is neither fully successful nor failed. The user has to wait for the system’s data to be fully synchronized before getting the final status of whether or not the ticket was purchased successfully.

BASE theory recognizes that in a distributed system, the state may soften over time rather than reaching a consistent state immediately**.

This means that we need to tolerate some state uncertainty, e.g., we are not sure if we can waitlist in a train ticket waiting queue.

5.3 Final Consistency

Eventual consistency is the core idea of BASE theory. It states that ** distributed systems can remain inconsistent for a period of time, but eventually converge to a consistent state. **

It is not like strong consistency, which requires partitioned data to be consistent in real time, resulting in a costly synchronization of system data. It is also not like weak consistency, data updates are not guaranteed to be consistent, resulting in subsequent requests can only access the old data.

Most of the current industry’s distributed systems, and even relational database systems, are realized with eventual consistency. For example, MySQL’s master-slave backups use binlog logs and listening threads to keep the data in the slaves and masters eventually consistent over time.

In a nutshell, the BASE theory actually sacrifices strong consistency of data across nodes and allows data from different nodes to be inconsistent over time to obtain higher performance and high availability.

In a monolithic system, databases can still achieve strong consistency of transactions through ACID, but distributed transactions need to consider node communication delays and network failures.

Therefore, BASE theory is the scheme we often use in real distributed systems.

​Market Overview: A Robust Recovery Amid Policy Shifts​​

The U.S. stock market witnessed a dramatic rebound in late May 2025, driven by easing trade tensions and resilient tech sector performance. On May 27, the S&P 500 surged 2.05% to 5,921.54, nearing its February all-time high, while the Nasdaq Composite jumped 2.47% to 19,199.16, marking its strongest single-day gain since April. The Dow Jones Industrial Average rose 1.78%, closing at 42,343.65, as investor sentiment shifted from caution to optimism. This rally followed weeks of volatility tied to trade policy uncertainty, with the S&P 500 now up 6.3% year-to-date and the Nasdaq reclaiming a 10% gain for 2025.

The turnaround was catalyzed by President Trump’s decision to delay imposing 50% tariffs on EU goods until July 9, coupled with progress in U.S.-EU trade negotiations. This move alleviated fears of an immediate economic shock, with Apollo Wealth Management’s Eric Steiner noting, “The market is learning to navigate Trump’s negotiation tactics, pricing in pauses but remaining wary of abrupt reversals”.

​​Tech Stocks Lead the Charge: AI and Semiconductor Dominance​​

Technology giants spearheaded the recovery, reflecting their outsized influence on market momentum. Tesla surged nearly 7% on renewed enthusiasm for electric vehicles and AI-driven manufacturing innovations, while Nvidia climbed over 3% ahead of its earnings report, which projected a 66% year-over-year revenue jump. The “Magnificent Seven” (Apple, Microsoft, Amazon, Meta, Alphabet, Nvidia, and Tesla) collectively added $320 billion in market cap, underscoring their role as market stabilizers.

Semiconductor stocks also shone, with KLA Corp and ASML rising over 4% and 3%, respectively, fueled by demand for AI infrastructure and easing supply chain constraints. The sector’s resilience aligns with Fidelity International’s 2025 outlook, which emphasized tech earnings as a “profit pivot” for broader market gains.

​Trade Policy: A Double-Edged Sword​​

Tariff developments remained a critical driver. The May 27 tariff delay reversed a mid-May selloff triggered by Trump’s earlier threats, which had sent the S&P 500 down 2.6% in a single week. Investors now appear desensitized to trade rhetoric, with Paul Nolte of Murphy & Sylvest Wealth Management observing, “The market has decoded Trump’s brinkmanship—sell on threats, buy on delays”.

However, risks linger. Citi’s Chief Economist Nathan Sheets warned that Trump’s “unprecedentedly aggressive” trade policies could slash global growth to 2.3% in 2025, with U.S. deficits projected to average 6% of GDP. The World Bank echoed concerns, noting that prolonged tariffs might shave 0.5% off U.S. GDP by year-end.

​Economic Data: Mixed Signals and Inflation Anxiety​​

Recent economic indicators painted a nuanced picture. The May Consumer Confidence Index rebounded 14.4%, signaling renewed household optimism, while the U-3 unemployment rate held steady at 4.1%. However, the U-6 underemployment rate ticked higher, and 44% of consumers anticipated rising joblessness—a red flag for spending.

Inflation remains a wildcard. Despite core CPI cooling to 3.2%, Trump’s “reciprocal tariffs” have pushed 1-year inflation expectations to 7.3%, the highest since the 1980s. Atlanta Fed President Raphael Bostic reiterated concerns, stating, “Tariffs could reignite price pressures, complicating the Fed’s path”.
​

​Sectoral Performance: Divergence and Opportunity​​

​1. ​Winners: Tech and Consumer Discretionary​​
- The S&P 500’s tech and consumer discretionary sectors soared 3.1% and 2.8%, respectively, on May 27. Companies like Shopify (+13.7% in mid-May) and Airbnb (+9%) capitalized on AI-driven efficiency gains and pent-up travel demand.
- Automotive stocks rallied, with Stellantis up 6% and Toyota gaining 2%, buoyed by tariff relief and EV subsidies.
2. ​​Laggards: Gold and Healthcare​​
- Gold miners slumped as risk appetite returned: Gold Resources plunged 6%, and Newmont Corp fell 4%.
- Healthcare underperformed, with the S&P 500 Health Care Index flatlining amid policy uncertainty and a 62% crash in Rocket Pharma after FDA halted its gene therapy trial.

​​China’s Tech Dilemma: Contrasting Fortunes in U.S. Listings​​

Chinese ADRs faced turbulence. While the Nasdaq Golden Dragon Index dipped 0.28%, outliers like Bilibili (+2%) and Tencent Music (+2%) defied the trend. However, Pinduoduo nosedived 13.6% after missing Q1 revenue estimates (956.7billionvs.985 billion expected) and reporting a 45% profit drop. EV makers also struggled, with Nio and XPeng down over 3% amid tariff-related supply chain fears.

​Looking Ahead: Navigating a Fragile Equilibrium​​

The market’s trajectory hinges on three factors:

1. ​​Fed Policy​​: With rate cuts paused, Chair Powell’s June remarks will clarify whether stubborn inflation warrants prolonged tightening.
   ​​2. Earnings Momentum​​: Q2 reports, especially from AI-centric firms like Nvidia and AMD, must validate lofty valuations.
   ​3. ​Geopolitical Risks​​: Escalating Middle East tensions and U.S.-China tech decoupling could reignite volatility.

Goldman Sachs strategists advise a barbell approach: “Overweight AI innovators and defensive utilities, while hedging with Treasury futures against policy shocks”.

​​Conclusion: Cautious Optimism in a Policy-Driven Market​​

May 2025 underscored the U.S. stock market’s resilience amid whiplash-inducing headlines. While tech strength and delayed tariffs fueled gains, high valuations (S&P 500 P/E ratio: 38.5x) and political unpredictability demand vigilance. As Fundstrat’s Tom Lee cautioned, “The rally’s second act requires earnings to catch up with prices—or risk a 10% correction”. For now, investors ride the updraft, but with seatbelts fastened.

2025 has emerged as a pivotal year for quantum technology, with both academic and commercial milestones reshaping the field. China’s University of Science and Technology (USTC) unveiled the ​​Zuchongzhi-3​​ superconducting quantum computer, boasting 105 functional qubits and outperforming classical supercomputers by ​​15 orders of magnitude​​ in solving quantum random circuit sampling tasks. This achievement, published in Physical Review Letters, solidifies China’s position in the global quantum race, closely rivaling U.S. advancements like Google’s latest Sycamore iterations.

Meanwhile, ​​D-Wave​​ launched its ​​Advantage2​​ quantum annealer, designed for optimization and AI tasks, featuring a 40% boost in energy efficiency and 75% noise reduction. In a parallel breakthrough, Australian researchers at the University of Sydney demonstrated ​​single-ion quantum simulation​​ of organic molecular dynamics, a leap toward practical quantum chemistry tools. Such innovations highlight quantum computing’s transition from theoretical promise to real-world utility, with applications in drug discovery, logistics optimization, and climate modeling.

​​Artificial Intelligence: Smaller Models, Bigger Impacts​​

The AI landscape in 2025 is marked by the rise of ​​compact, energy-efficient models​​ that challenge the dominance of large language models (LLMs). Chinese firm ​​DeepSeek​​ disrupted the market with ​​R1​​, an open-source model achieving performance comparable to OpenAI’s GPT-4 at a fraction of the cost ($5.57 million vs. billions). R1’s ability to run locally on edge devices enhances privacy and democratizes access, empowering schools and SMEs.

Generative AI has also evolved. ​​Sora​​, OpenAI’s text-to-video model, now generates hyper-realistic 3D environments, while ​​Huawei​​ and ​​China Mobile​​ showcased ​​5G-A-enabled humanoid robots​​ performing complex tasks like precision welding and disaster response. Ethical governance remains critical, as China’s AI Technology for Good White Paper sets global standards for data security and algorithmic transparency.

Biotech: CRISPR Expands Its Healing Horizon

Gene editing continues its therapeutic revolution. Following the 2023 approval of ​​Casgevy​​ for sickle cell disease, 2025 sees ​​CRISPR-based therapies​​ targeting chronic hepatitis B, age-related macular degeneration, and autoimmune disorders. Researchers are now engineering ​​universal CAR-T cells​​ using CRISPR, enabling off-the-shelf cancer immunotherapy and reducing treatment costs by 50%.

In HIV prevention, a ​​biannual injectable drug​​ reported ​​0% infection rates​​ in trials, offering hope for eradicating the virus in underserved regions. Meanwhile, ​​AlphaFold3​​ has mapped over 200 million protein structures, accelerating drug design and pandemic preparedness.

​​Space Exploration: A Crowded Moon and Cosmic Insights​​

The Moon is buzzing with activity. Japan’s ​​iSpace​​ and U.S.’s ​​Intuitive Machines​​ are deploying landers to the lunar south pole, searching for water ice to support future colonies. China’s ​​Tianwen-2​​ mission, set for mid-2025, aims to return samples from asteroid ​​2016 HO3​​ and study comet ​​311P​​, advancing planetary defense strategies.

Astrophysics leaps forward with NASA’s ​​SPHEREx​​ telescope, launching in February to survey 450 million galaxies and 100 million stars, unraveling cosmic mysteries like dark matter. The ​​Solar Wind-Magnetosphere Imager (SMILE)​​, a Sino-European satellite, will visualize solar wind interactions with Earth’s magnetic field, improving space weather forecasts.

​​Green Tech: Energy Innovation Meets Climate Urgency​​

Nuclear energy is reborn. ​​Small Modular Reactors (SMRs)​​ are gaining traction, with tech giants like Google and Amazon investing in nuclear-powered data centers to meet AI’s energy demands. China’s ​​Hualong One​​ reactor and the ​​BZ26-6​​ offshore oilfield—the world’s largest metamorphic rock reserve—underscore its energy diversification.

In agriculture, ​​methane-reducing feed additives​​ for cattle have slashed emissions by 30%, while ​​sustainable aviation fuels​​ derived from waste oils and CO₂ are scaling globally. The ​​COP30​​ summit in Brazil will test nations’ resolve to fund these transitions.

​​Neurotechnology and Robotics: Merging Mind and Machine​​

China’s ​​NEO​​ brain-computer interface (BCI), rivaling Neuralink, enables paralyzed patients to control robotic limbs with 95% accuracy. Meanwhile, ​​autonomous taxis​​ operate in over a dozen cities, with Baidu and Tesla vying for dominance in AI-driven mobility.

Robotics sees a paradigm shift. ​​General-purpose robots​​ trained via generative AI now adapt to dynamic environments, from factory floors to disaster zones, reducing deployment times from months to hours.

​​Conclusion: A Global Tech Race Redefined​​

As 2025 unfolds, the fusion of ​​quantum computing​​, ​​AI ethics​​, and ​​biotech breakthroughs​​ underscores a collaborative yet competitive global ecosystem. China’s strides in nuclear fusion, quantum networks, and green infrastructure signal its ambition to lead, while U.S. and EU innovations in AI governance and space exploration highlight diverse priorities. The challenge lies in ensuring equitable access to these technologies, lest the digital divide deepen into a chasm.

However, some of our friends who have just immigrated to the United States are not very clear about what taxes are exempted and why they are exempted.

Today, let’s talk about what taxes can be exempted from IUL policies in the United States!

Life Insurance Death Benefits

A death benefit from life insurance is a benefit from the estate, not subject to personal income tax, and subject to estate tax.

So why doesn’t life insurance count as income?

An example:

Suppose you own a house and have homeowners insurance.

If you are unfortunate enough to have your house destroyed by fire one day and the insurance company pays you the premium, should this premium be counted as your income? Of course not! Because you don’t have a house anymore, the insurance money is the money that compensates for the loss of the house.

The same applies to life insurance. Assuming that the insured is the main source of income for the family, if something happens to him, the death benefit is actually a replacement for him to continue to contribute to the family financially.

This is the logic behind the fact that the death benefit of a life insurance policy is not counted as income.

So how do you pay estate tax on a death benefit from a life insurance policy?

There are two scenarios:

1, Life insurance purchased by foreigners: Regardless of the status/nationality of the beneficiaries, no estate tax is paid on the death benefit as long as the policyholder is a non-U.S. person. Foreigners who purchase real estate in the U.S. are subject to U.S. estate tax in the event of an accidental death, and the exemption is not $11.4 million, but $60,000 instead.

2, life insurance purchased by U.S. persons: the amount of death benefits received by the beneficiary plus all other inheritance, in the estate tax exemption within the exemption amount of the tax-free, exceeding the part of the estate tax is required to pay the estate tax, this exemption amount is often changing, in 2019 for the 11.4 million U.S. dollars. The excess is taxed at an excessively progressive rate, such as 40% of the tax rate on the excess of $1 million.

Cash appreciation in life insurance accounts

Cash value appreciation in a life insurance account is exempt from personal income tax, but certain conditions must be met.

Life insurance premiums are subject to individual income tax before they are deposited. The reason that the cash value appreciation of the premiums after they are deposited into the insurance policy is tax-free is because most people use their money in the form of principal withdrawals or loans. Where the principal is taxed money, the loan is not income, so neither is subject to income tax.

When some brokers sell life insurance, some key details are not made clear, such as this point, intentionally or unintentionally vague as “withdrawals from life insurance policies are tax-free”.

You can only withdraw money from a policy that has a savings and investment function. The amount of money that can be withdrawn is determined by the cash value of the policy.

Why emphasize policy loans?

Because the death benefit of a life insurance policy is originally left to the beneficiary after the person passes away, but now we can use it for ourselves, in fact, we are actually “borrowing” the death benefit from the beneficiary of our policy in advance to use it. The collateral for the “loan” is the cash value of the policy. This is why the amount of money that can be “borrowed” from the policy must be within the cash value of the policy.

Since it’s a loan, naturally there will be interest. But on the one hand, the money is your own, and on the other hand, since the insurance company wants to make “getting money while you are alive” as a selling point, the interest rate charged is extremely low, and in some cases, it is even zero percent.

Prior to 1980, the money put into life insurance was tax-free and no conditions were set.

Later, more and more people used insurance to avoid taxes, so exaggerated that some people actually put millions of cash into the policy at one time, but the insured amount is only a few tens of thousands of dollars, and the insured used the policy to avoid a large amount of personal income tax.

In order to cope with this situation, the IRS issued several bills after 1984, stipulating what kind of life insurance policy money is tax-free, these regulations are still in effect.

For a policy to be tax-free, two conditions need to be met:

1, the amount of annual premiums paid for the policy cannot exceed the maximum amount set by the IRS, otherwise it will need to be taken out in the future and taxed on the gross proceeds.

The IRS has Life Insurance Definitional Testing and MEC (modified endowment contract) testing for every life insurance policy, and those that pass the test enjoy tax exemption.

What happens if a policy fails the test and becomes a MEC policy?

a. Policy withdrawals are taxed by the IRS on the basis of the total appreciation in value beyond the principal amount of the policy, and the tax is calculated on the basis of the individual income tax.

b. Money withdrawn from the policy before age 59½ is subject to a 10% penalty to the IRS.

c.Once a policy is recognized as a MEC, it is a MEC for life.

d.At the time of premium payment, the policyholder has two months to adjust the payment if he or she finds that the payment amount exceeds the maximum amount that can be saved for the year. Nowadays, every policy will do this test for the customer at the proposal stage, and the broker will clearly inform the customer of the maximum amount that can be saved in a year, the recommended number of years to save, how much to save per year, what will happen if it exceeds, etc.

2, The policy cannot be terminated prior to death at the time of the drawdown.

If the policy is terminated before death, the IRS will consider the policy not as insurance, but as an investment.

The investment is subject to capital gains tax, which is levied on the total amount of money withdrawn by the client minus the total amount of principal invested. The tax rate is based on the long-term capital gains tax rate and is related to the tax bracket in which the insurance policy is placed in the year of termination.

Life insurance termination (Lapse), we often see this word in a plan.

For example, if you save $10,000 per year for 20 years and save a total of $200,000, and you haven’t had a chance to withdraw the money yet, and then in the 25th year, the policy Lapses, then you don’t have to pay income tax. But if you withdraw money from the policy, say for 10 years at $50,000 a year for a total of $500,000, and the policy Lapses in the 30th year, then $500 - $200 = $300,000 is subject to capital gains tax.

Premium tax paid by insurance companies to their customers

When premiums are paid, the insurance company pays a premium tax on behalf of the customer. All life insurance companies charge a fee, usually 6-8%, for each premium received, and this amount includes the premium tax charged by the states, which is about 3%.

This portion of the tax is mandatory and is paid on all life policies in the U.S. Tax rates vary slightly from state to state. Each company charges a slightly different rate and collects it in a slightly different way. Now there is some understanding of life insurance tax exemption in the United States Boo!

1. basically determine which insurance company you want to buy insurance from before you go to the dealer to finally buy the car;
2. go to the dealer to look at the car and decide on the price and exactly which car;
3. the dealer will handle all the transaction formalities, including registration and temporary license plate, at which time you are required to provide insurance;
4. at this time you call the insurance company directly at the dealer, or log on to the official website of the insurance company to buy insurance, provide your driver’s license information, vehicle information, and the insurance options you need, the insurance company will send the proof of insurance to the dealer within a few minutes;
5. The dealer receives the certificate, and after completing the procedures mentioned in 3, you can drive the car away.

So how do you pick an insurance company?

The larger insurance companies in the US like GEICO, Progressive, Allstate, etc. Go directly to their website to get a quote. The quote only needs general information about the vehicle, such as the model number, but not the VIN number, so you can get a general idea of the price before you go to the dealer. For these big insurance companies, there is basically no difference in service, so just pick a relatively cheap one.

If you already have other insurance with a particular insurer, such as homeowner’s insurance, etc., you can prioritize that insurer as there are usually some discounts. However, it is still advisable to inquire about a few more.

If one is a member of AAA or costco, one can consider the insurance they are associated with as there are some discounts as well.

What kind of insurance do you choose?

Similar to domestic car insurance, American car insurance also has many options, basically divided into two Liability and Comprehensive.

Liability is similar to the domestic compulsory insurance, is the law requires the car on the road to purchase compulsory insurance, generally only includes the accident in the case of your fault on the other side of the compensation.

Comprehensive can be called all-risk insurance, which is to compensate both the other party’s damage and your own.

When choosing an insurance policy there are two parameters for each program that directly affect the price:

1. Limit, which is the amount of insurance coverage, is the maximum amount of compensation that the insurance company will pay in a single accident.

2. Deductible, which is the amount of deductible or co-payment, is the portion of a single claim that is to be borne by the individual, beyond which the insurance company will pay.

The choice of insurance and the amount depends on individual needs. For new cars, it is still recommended to choose full coverage insurance, limit choose the minimum is good, deductible can choose $1000.

In addition, if you are buying a car with a mortgage, the lender usually has higher requirements for car insurance than compulsory insurance, and the dealer will tell you before you buy.

Distributed Ledger Technology, the definition

Distributed ledger technology is a digital system that records the transaction of assets by recording their details in multiple places at the same time. When compared to traditional databases, distributed ledgers lack a central data store or administration functionality. Distributed ledger technology refers particularly to the technological composition and protocols that facilitate the simultaneous access, validation, and keeping of records up to date which is the main characteristic of distributed ledgers. It functions of a computer network distributed over different multiple entities, locations, or nodes. A distributed ledger consists of nodes that process and verify every item, thereby creating a record of each item and developing a consensus on its accuracy. A distributed ledger can record different types of static and dynamic data including a registry and financial transactions. Blockchain is a typical example of a distributed ledger technology.

How Does Distributed Ledger Technology Work?

distributed ledger technology focuses on the principles of decentralization. In contrast to conventional centralized databases, distributed ledger technology works on a peer-to-peer (P2P) network, where several nodes store, validate, and update the ledger at the same time. This removes the burden of a central authority and lessens the chance of a single point of failure. The process is accomplished by replicating the digital data across the network of nodes. Each node maintains an identical copy of the ledger and single-handedly processes new update transactions. To ensure consensus, all participating nodes employ a consensus algorithm that determines the correct version of the ledger. Once harmony is accomplished, the updated ledger is spread to all nodes, which ensures synchronization and efficiency. Distributed ledger technology uses cryptography to safely store data and cryptographic signatures and keys that only authorized users can access. The technology also develops a fixed database, which means the stored information cannot be erased and all updates are permanently recorded for the future.

This buildup represents an important difference in how information is gathered and communicated. It does this by moving record-keeping from an individual, authoritative location to a decentralized system in which the invited parties can view and modify the ledger. Therefore, different parties can see who uses and modifies the ledger. This is a very transparent feature of distributed ledger technology, one that provides a high level of trust among the participants and practically removes the possibility of fraudulent activities happening in the ledger. Therefore, distributed ledger technology eliminates the burden of entities using the ledger to depend on one central authority that controls the ledger or a third-party provider to handle the task of checking against manipulation.

Industries that Use Distributed Ledger Technology

The use of distributed ledger technology goes across a wide range of industries and is changing conventional processes. Common industries that use this technology include the following.

Banking and Finance

The banking and finance industry is a proponent of distributed ledger technology especially in the implementation of smart contracts in trade finance. Smart contracts ease the proper functioning and settlement of trade transactions, decreasing lags and removing the need for middlemen. Also, distributed ledger technology ensures quicker cross-border payments, enhances processes, and offers secure digital identity solutions.

Supply chain management

Distributed ledger technology is also an aspect of supply chain management. Distributed ledgers help companies to monitor and validate the movement of goods, with authenticity and no fraud. This technology facilitates real-time visibility into supply chain operations, decreases paperwork, and reduces lags. For example, companies can use a distributed ledger solution to track the origin of goods, in an ethical and trustworthy way.

Healthcare

Distributed ledger technology is also useful in the healthcare industry to improve patient data management, streamline processes, and enhance security. This technology can facilitate the storing of medical records, ensuring data privacy and integrity. Also, smart contracts can automate insurance claims, minimizing administrative burdens and improving efficiency. Distributed ledger technology also makes it easy to conduct safe and transparent clinical trials, ensuring the integrity of data and building trust in the research process.

Real estate

The use of Distributed ledger technology can help improve the real estate industry by facilitating property transactions, decreasing paperwork, and enhancing security. Since it deals with smart contracts, companies can automate property transfers accurately providing a tamper-proof record of ownership. Blockchain platforms built on distributed ledgers are transparent, reducing the risk of fraud and disputes. It also removes the need for costly middlemen. Furthermore, this technology ensures partial ownership of real estate, creating new venture opportunities and adding liquidity to the market.

What the future holds for Distributed Ledger Technology

With the resounding progress of distributed ledger technology, many still ask whether it can revolutionize how governments, institutions, and industries work. Many experts are promoting distributed ledger technology as an important asset that will not only revolutionize the existing processes but could soar innovative new applications. Moreover, distributed ledger technology is seen as part of the “internet of value,” where transactions occur in real-time across global networks. The reason why digital ledger technology exists is because of the pervasive internet that enables it. Nonetheless, according to experts, the adoption of this technology will follow the typical technology curve, led by a few exemplary leaders, then fast followers, and, finally, the laggards. Experts also believe that it won’t be easy for organizations to implement, scale, and operate distributed ledger technology. To that extent, enterprise executives and entrepreneurs are now faced with the burden of establishing networks that can take advantage of distributed ledger technology to revolutionize how they share and keep records.

I. What is semi-insurance and full insurance?

Semi, also known as Liability Insurance, is the most basic form of automobile insurance. It is used to pay for property damage and bodily injury costs of the other party in an automobile accident. It’s important to note that this insurance typically does not cover the cost of your property and personal injury damages. Specifically, semi-insurance includes:

1. Property Damage Liability: If you are found to be at fault in an accident, this insurance policy will cover the cost of repairing or replacing the other party’s property, including vehicles, fences, buildings, and signals, according to the terms of the policy.
2. Bodily Injury Liability (BIL): If you are found to be at fault in an accident, this insurance policy will pay for the medical expenses, rehabilitation costs, and other related medical expenses of any person injured in the other vehicle, according to the terms of the policy.

The Minimum Liability Limit varies from state to state in the U.S., but is generally between $10,000 and $50,000.

Take New York State, for example:

• Minimum $10,000 for property damage caused by a single accident.

• 25,000 for single person bodily injury and $50,000 for death.

• Minimum $50,000 for multiple bodily injury and $100,000 for death.

It is important to note that insurance companies generally pay a maximum of $100,000 to $500,000 for liability coverage. If an accident occurs and the maximum insurance coverage is not enough to pay the full compensation, you will have to pay the difference out of your own pocket. Therefore, Tiger suggests you to decide the maximum benefit amount according to your own property.

Full Coverage Insurance is a more comprehensive type of auto insurance. It is used to pay for property damage and bodily injury costs for you and the other party after a vehicle accident. Specifically, full coverage insurance must be included:

1. Liability Insurance: If you are found to be at fault in an accident, this insurance will pay for the cost of repairing or replacing the other party’s property and related medical expenses, according to the terms of the policy.
2. Collision Coverage: If your vehicle is damaged or destroyed in an accident with another vehicle, regardless of who is at fault, this coverage will pay for the cost of repairing or replacing the vehicle for both you and the other party, according to the terms of the policy.
3. Comprehensive Coverage: This coverage is designed to protect your vehicle against a wide range of non-collision damages or risks, with claims paid according to specific terms. Common causes of claims include fire, theft, natural disasters, falling objects, animal collisions, and malicious damage by others.

In addition to these three, car owners can also choose their own additional insurance. Common additional insurances include:

Medical Payments Coverage: This coverage will pay for medical related expenses incurred by you and your passengers in a car accident, such as ambulance fees, surgical procedures, prescription drugs, medical imaging, funeral expenses, etc., according to the terms of the policy. Medical payments insurance provides compensation regardless of who is responsible for the accident. Typically with this coverage, the insurance company will pay for the medical expenses directly without you having to pay a deductible or make an advance payment.

Uninsured/Underinsured Motorist Coverage: This coverage is used to protect against a vehicle accident where the other driver is at fault but is uninsured or underinsured.

Personal Injury Protection (PIP): This coverage pays for medical, lost wages, funeral and other expenses incurred by you and your passengers in a car accident, regardless of who is responsible for the accident. In the United States, sixteen states are mandatory for car owners to purchase this insurance, Tiger suggests that you consult the specific requirements in time when shopping for this insurance content.

Rental Reimbursement: This insurance is used to compensate for the cost of renting a car after an accident due to the need to repair the car, regardless of who is responsible for the accident. This type of insurance usually has a daily or per-accident maximum limit. You can choose different limits when purchasing your insurance, depending on your needs and budget.

Emergency Road Assistance Service (ERAS): This insurance is used in the event that your vehicle breaks down or becomes inoperable, the insurance company will pay for emergency services according to the terms and conditions, such as on-site repair and towing services for the owner of the broken down vehicle.

Note: All provisions of general liability insurance do not pay for the insured’s own damages. If the liability is on the other party, the other party’s liability insurance pays for you. If you are responsible, then your bodily injury and property damage will be covered by Medical Payments Coverage, Collision or Comprehensive, or your own medical insurance may be able to intervene (depending on your medical insurance). If you do not have auto insurance or medical insurance, you will have to pay for your own medical care.

II Which one should I choose, half or full coverage?

Choosing the right half and full coverage U.S. auto insurance for you depends on your personal needs, vehicle value, financial situation, and risk tolerance. Here are a few steps to help you make a better choice of the type of auto insurance you want:

Check Legal Requirements and Minimum Requirements: you need to know the legal requirements for auto insurance in your state. Each state has different legal requirements; some states only require car owners to purchase minimum liability insurance, while others require broader coverage auto insurance.

Check for Loan and Lease Requirements: If you have a loan or lease on your car, your lender or lessor will likely require you to purchase full coverage insurance. This is because they want to make sure the car is protected in the event of an accident.

Assessing the value of your vehicle: If your car is new or has a high value, it is more appropriate to purchase full coverage. This is because the cost of repairing or replacing your car in the event of an accident can be high. If your vehicle is more than ten years old, cheaper or used, semi-insurance may be more appropriate. This is because it will be cheaper to repair or replace car parts in the event of an accident.

Evaluate your personal situation and risk tolerance: If you are an experienced driver and can afford to pay for the damage to your vehicle, you can save a lot of money on your premiums by choosing semi-insurance. If you are new to driving or feel you cannot afford the costs associated with vehicle repair or replacement, full coverage may be more appropriate for you.

Frequency of Vehicle Use: If you drive frequently and drive in areas with poor road conditions, full coverage can provide more comprehensive protection. If you drive short distances and less frequently, and the roads in your driving area are in good condition, you will save a lot of money by choosing semi-permanent insurance.

Consult a professional: You can contact Tiger Tiger’s customer service to ask about auto insurance, Tiger Tiger will provide you with personalized advice and help you choose the right type of insurance and coverage.

If you want to learn more about auto insurance related content, check out our auto insurance blog posts. Tiger Tiger provides detailed information on the factors that affect the price of auto insurance , what to consider when buying auto insurance, ways to reduce the cost of auto insurance, and the US auto insurance vocabulary section!

Some of my friends may have had such an experience, buying insurance is too expensive and too troublesome, and then regretted that they had bought less when the insurance claim was made. In this era of ubiquitous risk, buying enough insurance is a reliable life protection for yourself and your family.

The five most important types of insurance to buy

Auto Insurance

Auto insurance is a no-brainer. Almost everywhere in the U.S., there are laws mandating that drivers must have insurance before they can hit the road. If you don’t have insurance, or if you don’t have enough insurance, you will have to pay for your own medical expenses, car repairs, and everything else in case of an accident. If you are responsible for your own injuries and cause someone else to suffer some serious permanent injuries, such as paralysis and disability, the amount of long-term medical expenses and other liability compensation is very high.

Don’t leave it to chance. According to statistics, there are an average of 6 million recorded car accidents in the U.S. each year, nearly 40,000 people die in car accidents, and about 2 million are permanently injured in car accidents. According to data released by the CDC in 2010, medical and compensation costs for car accident injuries and deaths in the U.S. totaled more than $99 billion in one year.
So if you want to save money on auto insurance, you can consider raising your deductible or buying only half coverage if you drive an older car, but the amount of liability insurance must not be reduced, and many professionals recommend purchasing more than the total value of your own net worth of total liability insurance coverage. Of course, driving a new car with spare money partners should also consider collision insurance and comprehensive accident insurance, etc., if their car crashed to be repaired, insurance can also be reimbursed to reduce out-of-pocket expenses.

Health Insurance

As the saying goes, “Health is the greatest asset in life,” and in the United States, health insurance is indispensable. Most regular health insurance policies allow you to get regular checkups and vaccinations for free or at a very low cost, and also reimburse you for most of your medical expenses when you are sick, so that you can afford to pay for them.
There was once a study that said the number one reason Americans go broke is medical expenses →_→

The United States medical cost is very expensive, see an ordinary family doctor will be two or three hundred knives, if a blood test or do a special examination, casually is a few hundred, to see a specialist is even more expensive; not to mention the cost of surgery and hospitalization, thousands are idle. The average working family is not far from bankruptcy if they get sick once.
So even though buying health insurance itself isn’t cheap, many people would rather have the most basic insurance than nothing at all. Many young people in the United States can’t afford regular medical insurance premiums, so they will buy a cheaper Short Term Health Insurance or Hospital Indemnity Insurance, at least in case of a serious illness, there will be insurance to help reimbursement.
Luckily, many employers also provide health insurance benefits, which can save the average working class a lot of insurance costs. Therefore, when looking for a job, in addition to the salary, you must also consider whether there are other benefits such as health insurance, which is also a hidden wealth that should not be underestimated.

Home Owner’s/Renter’s Insurance

In the U.S., if you have a loan to buy a house, the bank will force you to buy Home Owner’s Insurance. In case of fire or wind damage to the house, or if the house is stolen or damaged, or if the homeowner is sued and needs to compensate other people and so on, the homeowner’s insurance will pay for the homeowner’s losses, and it also protects the homeowner’s ability to continue to repay the loan.
Even if it is not mandatory, homeowner’s insurance is a wise investment. This is because in addition to covering the construction of the house, homeowner’s insurance usually covers claims for damage to the homeowner’s personal belongings. For example, if your car is smashed or your laptop is stolen from your car, the comprehensive accident insurance of your auto insurance will only cover the cost of repairing the car, but not the personal belongings such as the laptop, whereas homeowner’s insurance will cover such losses.
In addition, homeowner’s insurance usually includes homeowner’s liability insurance. For example, if someone falls and gets hurt on your property, homeowner’s insurance will reimburse the medical and related expenses; or if you are sued for damages to someone else’s property, homeowner’s insurance will pay for the costs of the lawsuit as well as the compensation.

However, homeowners should also be aware that general homeowners insurance does not cover damages caused by natural disasters such as floods, earthquakes, and landslides, so it is best to consider purchasing additional insurance if your home is located in a high-risk area where these types of natural disasters are common.
Although renters don’t have to worry about damage to their home, renters insurance will cover personal belongings and personal liability, as well as reimbursement for the cost of renting a hotel or other place to stay if your apartment is temporarily uninhabitable due to a fire, for example. Renters insurance in the United States is only $187/year on average, usually only $15~$30/month, less to go out to eat on the affordable, or very cost-effective.

Disability Insurance

There are unpredictable winds and clouds, if an accidental injury, serious illness and so on, resulting in not being able to go to work without a salary, it is likely to affect the normal life (including payment of rent, mortgage, utilities and other daily expenses, etc.), to buy disability insurance for when you lose the ability to work, the insurance can be another way to protect your income.
Although the U.S. social security system is relatively complete, Workers’ Compensation only insures work-related injuries, Unemployment regardless of the incapacity to work, apply for Social Security Disability Insurance (SSDI) generally need to wait for at least 5 months, and the approval is very strict and easy to be rejected, and even if the successful applicant has been rejected, it is easy to be rejected. Even if the application is successful, the average benefit in 2018 is $1,197/month, which is far from making up for the normal income.
Disability insurance is mainly divided into Short-Term (referred to as STDI) and Long-Term (referred to as LTDI) two categories, most of which will be paid out at about 60% of the pre-tax income, because individuals who buy this kind of insurance, the benefits are usually tax-free, so it’s almost the same as the money that actually comes to your hand after deducting taxes.

Short-term disability insurance generally pays out for a shorter period of time, about three months to a year, but can start receiving benefits sooner (usually after 1 to 14 days), while long-term disability insurance has a longer waiting period to start receiving benefits, usually after 90 days, but keeps paying out for a longer period of time, which can be several years or even until retirement. The cost of disability insurance is usually about 1% to 3% of your original income, depending on the terms of the policy, and usually waiting times, benefit percentages, and benefit periods are optional.
Many employers also offer disability insurance benefits, so guys might want to check with HR. However, it is important to note that if your employer purchases the insurance, or if you choose to purchase the insurance with pretax income, you will be required to pay taxes in the event that you need to apply for benefits. In addition, according to the Society for Human Resource Management (SHRM), there are five states in the U.S. that provide or require employers to provide short-term disability insurance: California, Hawaii, New Jersey, New York, and Rhode Island, so you can check the official website for details.

Life Insurance

Death is not cheap in the U.S. According to GoBankingRates.com, the average death necessity expense in the U.S. is $11,618, and the median funeral out-of-pocket expense (which includes the cost of viewing and cremation of the remains) is $7,360.If you are the primary source of income for your family, that touch wood says that in case of XXX, your family will likely be in a financially difficult If you are the main source of income for your family, then touch wood says that in case of XXX, your family will probably be in financial difficulties, and purchasing a life insurance policy will give your loved ones a protection.
There are many different types of life insurance, but they can be divided into two basic categories: Term Life Insurance and Whole Life Insurance.

The so-called term life insurance, that is, in the insurance contract specified period (Term, usually ranging from 5-30 years), to provide insurance protection for the insured, beyond the contract period, nothing; while whole life insurance is not specified in the contract period, will remain in force until the death of the insured, and whole life insurance is usually accompanied by a deposit investment function, can accumulate the cash value, you have the need to “borrow” the cash amount to use. You can “borrow” the cash amount when you need it.
Life insurance premiums depend on the age and health of the insured, and whole life insurance can be several times more expensive than term insurance for the same amount of coverage for the same person. However, term life insurance is usually affordable, so it’s worthwhile to purchase a term life insurance policy when you don’t have a lot of money to spare.

Especially if you are the breadwinner of your family and are making payments on your home and car, it is important that you have a life insurance policy that lasts long enough (e.g., to cover the entire term of your loan, or until retirement age) and has enough coverage to pay off all of your debts, as well as your family’s future living expenses.

Of course, if you win a big lottery, your family is very wealthy, and your family will not be under much financial stress, then life insurance may not be necessary.

Five of the worst insurance policies to buy

Flight Insurance

Flying can actually be considered one of the safest and most secure ways to travel, and it’s a bit of a waste to spend money on insurance for such rare accidents, and you’re already covered if you have life insurance.

Mortgage Protection Insurance

This type of insurance is similar to term life insurance, but claims are limited to paying off the mortgage, unless you’ve been denied coverage by the insurance company due to a health condition, in which case it would be more cost-effective to buy term life insurance outright.

Cancer/Disease Insurance

It’s much more worthwhile to spend the money on a good health insurance policy, and this kind of insurance that only covers certain diseases is quite unnecessary.

Credit Card Insurance

Although there are a lot of ID Theft issues out there right now, it generally doesn’t cost you much actual money when it comes to identity theft issues as long as you report it properly. Because federal law states that if a credit card is stolen an individual is liable for up to $50, and many banks have their own Zero Fraud Liability, it won’t even cost you a dime as long as you report it in a timely manner. And as far as repairing your credit goes, these are completely free to do on your own, it just takes some time mentally.

Child Life Insurance

As mentioned earlier, buying life insurance is mainly for people who are the breadwinners of the family to buy protection for their family members, and there’s really no need to buy insurance for small children who don’t make any money; even though whole life insurance has some investment preservation features, it’s return on investment can be a lot smaller than the return on accounts such as 529 Plans, IRAs, and so on. However, if your child is a starlet and contributes to your family’s income, it’s worth it to get an insurance policy for your child.

Why not a second handshake?

In the case of a quadratic handshake, where the client sends a SYN to the server and the server replies with a SYN-ACK to the client, the connection is established at that point.

In this case, if the first SYN request is delayed in the network and the connection is established after the client resends the SYN, then when the delayed SYN request reaches the server, the server assumes that it is a new connection request, and at that point the client pays no attention to the server’s response, causing the server to keep on waiting and wasting resources.

Why not four handshakes?

Four handshakes add additional latency and complexity, and a fourth handshake doesn’t provide any additional assurance that three handshakes have solved the problem.

Three handshakes are already able to confirm that the sending and receiving capabilities of both parties are normal, and further confirmations would only add a round trip time and reduce the efficiency of establishing a connection.

Three handshakes to establish a connection:

1. Client sends SYN: The client selects a random sequence number x to send a SYN message and enters the SYN_SENT state.
2. Server sends SYN-ACK): the server receives the SYN message, selects its own sequence number y, and sends a SYN-ACK message, and the server enters the SYN_RCVD state.
3. Client sends ACK: the client receives the SYN-ACK message, sends an ACK message, and then enters the ESTABLISHED state.

Four hand waves after the end of a data transmission:

TCP connection termination, on the other hand, requires four hand waves, this is because a TCP connection is full duplex, i.e. both sides of the communication can send and receive information at the same time. When terminating a connection, each direction needs to be closed individually, so four waves are required.

1. Client sends FIN: the client decides that the data has been sent and sends a FIN message.
2. Server ACK: the server receives this FIN message, sends an ACK message to acknowledge it, and enters the CLOSE_WAIT state.
3. SERVER SEND FIN: The server sends a FIN message when it is ready to close the connection.
4. Client ACK: When the client receives the FIN, it sends an ACK message and then enters the TIME_WAIT state. After a period of time to ensure that the server receives the ACK message, the client closes the connection.

Life example:

You can compare the three handshakes to a telephone conversation. When you dial a phone number, the other person answers (the first handshake), you start greeting each other to make sure the other person can hear you (the second handshake), and then you start a conversation (the third handshake).
If you greet each other only once, you may not be sure the other person actually heard you; if you greet each other multiple times, it’s redundant and inefficient.

At the end of the call, you hang up when you say “goodbye” (the first wave), and the other person hangs up when they say “goodbye” (the second wave), thus ensuring that both parties understand that the call is over.

At the end of a phone conversation:

You first say “Do you have anything else, no I’m going to hang up” (the first wave), waiting for the other party to respond, which is equivalent to sending a FIN packet.

The other party responds “Let me think about what else” (second wave), which is equivalent to the other party sending an ACK packet, but the other party may still have some things to deal with, so the call is not immediately ended.

After a while, the other party confirms, “OK, I have nothing more to say, so hang up” (third wave), then the other party sends a FIN packet.

You respond with “Got it, hang up” (the fourth wave of the hand), corresponding to the ACK packet, after which both parties can hang up and end the call.

Combining our three-phase framework at LLM as-Copilot, LLM as-Integrator, LLM as-Facilitator, and our internal analysis materials, I would roughly summarize them into 6 trends:

1. from single-role assistance to end-to-end assistance.
2. Knowledge Management for Assisted Decision Making.
3. DevOps facilities for AI applications.
4. online fault localization and problem solving.
5. influx of AI-assisted UI design. 6. code translation vs.
6. code translation and inter-system translation.

Some of this knowledge is almost as much as we’ve previously reached agreement on, so let’s tell the story in reverse.

Digitalization of R&D forced by generative AI

Before we start the new trend summary, one thing we have to mention is: the digitization of R&D. Over the past year, I’ve talked to almost 10 companies’ R&D leaders about AI-assisted R&D. In fact, the reason that prevents most companies from applying generative AI, in addition to model limitations, is the poor level of digitization of R&D.

The first problem we have to face is that standardization is not on the ground. Simply put, standardization, platform, indicator-driven four maturity levels to consider the problem, some organizations are still in the standardization of the difficult problem of landing, not to mention indicator-driven improvement. Fortunately generative AI combined with tools can improve the problem of hard to land on specification, which is a potential opportunity to bend the curve - provided there is enough gumption to push forward.

In addition to this, the second problem we have to face is: Management of Knowledge - there is a lot of unspeakable knowledge in the organization (askew, e.g. content gossip). The challenges we will encounter are:

• Undocumented, unmanifested.
• Lots of outdated knowledge - you don’t know which document is old.
• A lot of non-textual knowledge - a conference whiteboard shot one day where the words are unrecognizable.

Simply put, these are part of our knowledge debt.

Code translation and inter-system translation

Scenario 1: Legacy System Migration. Generative AI features are well expressed in natural language translation and very prominent in programming languages. So, last year we also did the analysis of related features in AutoDev and built a series of related legacy system functions. In commercial products, we can also see such as IBM watsonx Code Assistant for Z such as Cobol to Java specialized tools.

And how to analyze the legacy system migration remains a complex issue. Existing tools are more often designed by people to migrate, assisted by AI.

Scenario 2: Inter-system translation. As, more and more large manufacturers start to develop Hongmeng applications, we have found in practice the advantages of generative AI in this regard. Since the UI differences of mobile systems are not that big, some of the functions can be migrated through translation. Although, we encountered a large number of generative AIs that lacked new proprietary knowledge (ArkUI, ArkTS, HarmonyOS APIs), a combination of combining Thought Chain and RAG with them could achieve more acceptable results.

The Emergence of AI-Assisted UI Design

AI generative code needs to incorporate information such as existing specifications in order to generate code that works. For back-end code development, where Spring reigns supreme, building this generative AI-friendly architecture is an easy task. However, generative AI is challenging in the front-end space as small, medium, and large organizations have their own brand guidelines, style guides, and design systems.

From the existing models, the main AI-assisted UI designs can be categorized into three types:

1. prototype generation to assist in demand communication.
2. UI design generation combined with low-code platforms. 3.
3. UI code generation combined with IDE plug-ins.

Considering the complexity of front-end requirements, it is obvious that it would be easier to start from the second scenario, while scenario 3 is more suitable for novices to learn and use the framework, and developers to use new frameworks.

Online fault localization and problem solving

Online Issue Fix. Before generative AI, conventional deterministic AI has enabled a great deal of automation. Conventional application performance monitoring (APM) tools can map errors reported by online runtimes to the corresponding faulty code. ps: Combined with information about the correlation between the requirements and the code, we can accurately deduce which requirement change caused the impact. With generative AI, online problems can be converted to fixing PRs to assist you in fixing the problem, such as NewRelic, which has a similar feature.

Fault localization. Troubleshooting networks and problems becomes incredibly important in complex systems that contain a large number of subsystems, such as a single microservice. In the absence of tools, humans often lose critical information at some point, and AI can assist us in solving such problems, such as AWS’s AI-assisted network troubleshooting.

Considering I’m an expert in Dev, not Ops, I can’t read much more into this.

DevOps facilities for AI applications

There are already a large number of online applications that have introduced AI capabilities, such as the Starbucks face-swap campaign, and these types of AI applications have introduced a series of AI infrastructures. Therefore, for medium to large organizations, in addition to considering the appropriate private deployment model, you also need to build a rapid AI DevOps infrastructure to support it.

In addition to the various types of monitoring of the big models themselves, we also need the operational costs of the models themselves – especially after you call third-party APIs to build a better AiBizDevFinGitSecOps system (🐶🐶🐶🐶). Naturally, we need an AI to advise you on your AI + Finance, such as building caching mechanisms, Prompt length optimization, and so on.

Knowledge Management for Decision Making

Knowledge management used to be a headache, now it has become a full-body pain (for lack of a better word). I’m sure you readers already understand generative AI very well:

• If you do not give him enough information, it generates the results can be accepted by luck.
• If you give him enough information, it will always ignore some important information to make you angry.

Angry or not, when you start thinking about landing, you start assuming that when I have an architectural specification, generative AI can assist will make architectural decisions. Then, you realize that you can’t find an architectural specification that meets the requirements. Similarly, in other scenarios, there are similar problems.

PS (a twist): So, you should consider prioritizing knowledge management as well, so that when you report back with your leaders, you can reasonably dump the pot.

From single-role assistance to end-to-end assistance

In fact, most of the above is about how AI can switch from single-role assistance to end-to-end assistance, just with requirements from different scenarios.

The difficulty with end-to-end assistance is not the design of the tool or the prompt itself, but whether the processes and specifications are in place. If there are problems with the processes and specifications themselves, then we need to explore whether there are more appropriate strategies from different scenarios.

Other and AI summaries

Of course, there are AI-assisted R&D scenarios such as instant assisted problem fixing that we’ve discussed in the past.

This article looks at trends in AI-assisted R&D in 2024, with a particular emphasis on the evolution of AI technology from simply assisting a single role to end-to-end assistance. The author begins by mentioning the importance of R&D digitization in AI adoption and points out the challenges of standardization and knowledge management. He then details six major trends:

1. From single-role assistance to end-to-end assistance: the AI technology is no longer limited to single-role assistance, but extends to all aspects of the entire R&D process.
2. Knowledge Management for Decision Aid: the application of AI in knowledge management has become more important, but also faces the problem of incomplete information and information selection.
3. DevOps facilities for AI applications: The introduction of AI applications requires the establishment of adaptable DevOps infrastructure to support their operation and monitoring.
4. Online Fault Location and Problem Solving: The application of AI in online fault location and problem solving is also becoming mature, which can help quickly locate problems and provide solutions.
5. Emergence of AI-assisted UI design: The application of AI in UI design has taken various forms, including assisting in demand communication, UI design generation on low-code platforms, and UI code generation by IDE plug-ins.
6. Code Translation and Inter-System Translation: The application of AI in code translation and inter-system translation is gradually maturing, especially in the performance of legacy system migration and inter-system function migration.

The announcement was quite harshly worded: “Mr Altman’s departure is the result of a considered decision by the Board of Directors, which believes that Mr Altman has not been sufficiently forthcoming in his communications with the Board of Directors and has impeded the Board’s ability to fulfil its responsibilities. The Board no longer believes he is capable of continuing to lead OpenAI.”

From a post by Greg, OpenAI’s chairman and president, the firing went like this: on Thursday night, Altman received a text message from Ilya, the chief scientist, about a meeting on Friday at noon. At noon on Friday, Altman accessed the meeting, the rest of the board, except Greg, were there, and Altman was told he was fired. At 12:19 p.m. Greg receives a call from Ilya, and at 12:23 p.m. Greg accesses the meeting and is told that he has been removed from the board of directors and continues to stay in management, and that Altman has been fired.

At the end of November last year, with the shiny introduction of ChatGPT, a wave of artificial intelligence swept the world, and Altman was honoured as the “father of ChatGPT”. A year later, he was fired, quite dramatic.

On November 6, Altman announced a series of product updates to developers and ChatGPT users around the world at OpenAI’s inaugural developer conference, including the ability to quickly create custom ChatGPT GPTs to realise everyone’s dream of owning a big model, and said he would compete with the likes of Microsoft and Google.

On Thursday, Altman also spoke at APEC, expressing bullishness on AI technology, arguing that the technology doesn’t need major regulation in the short term, and even claiming, “As a species, humanity is now on a path to self-destruction, and AI may be the means to stop it.”

And then suddenly he was “out of office” on Friday.

At 5:46 a.m. on November 18, Altman responded with a message on the X platform, “I loved my time at OpenAI, it changed me and hopefully it will change the world. Most importantly I have worked with so many talented people. I’ll explain more about what’s next later.”

OpenAI CTO Mira Murati, formerly CTO, has been named interim CEO.At the same time, company president Greg Brockman announced that he would be stepping down as chairman of the board of directors, but that he would continue to hold a position at the company, reporting directly to the CEO.

The board described interim CEO Mulati as technically well-rounded and particularly experienced in AI management and policy implementation. Previous reports have called her “a dynamic and conciliatory figure in AI” who has welcomed government regulation.

On Friday, she told employees at a company-wide meeting that Microsoft’s relationship with OpenAI remains stable following Altman’s departure. By the news of Altman’s departure, Microsoft shares late in the session once “diving” fell more than 2%, closing down 1.68%. Microsoft issued a statement that the company has a long-term relationship with OpenAI, abide by the investment commitment to OpenAI.

As you can see, Altman was ousted from the board for at least two reasons:

• One is that he disagrees with the board on the issue of how to handle the relationship with Microsoft.

• The second is a disagreement with the board on the question of whether regulation of AI should be prevented in the short term.

I. Begging for a package or independence?

Microsoft is by far the largest investment in OpenAI, after investing $1 billion in 2019, and another $10 billion in early 2023.

Microsoft and OpenAI were once called a “romantic alliance” in the tech world. However, if you look closely, you will find that the shrewd Microsoft has not lost its mind in the face of “romance”. 2019 investment of $1 billion in OpenAI, the conditions are very favourable to Microsoft, buy out the exclusive license of the basic technology of the GPT-3, most of the technology of OpenAI is preferentially licensed to the Microsoft products, and Microsoft has become the exclusive cloud provider of OpenAI, and Microsoft has become the exclusive cloud provider of OpenAI. OpenAI’s exclusive cloud provider, meaninga large portion of the amount of Microsoft’s investment that OpenAI received was to be paid to Microsoft as a cloud usage fee.

Last November, OpenAI released the ChatGPT chatbot, which took the world by storm within a few weeks. At the beginning of February this year, ChatGPT turned on a wild ride, and a new global AI race was triggered, with giants competing for the top spot. Microsoft immediately ChatGPT access to Microsoft cloud services, the launch of GPT-3, GPT-4 version of the Office Family Bucket and a series of new products, especially Microsoft ChatGPT into its search engine to launch a new version of the Bing artificial intelligence search engine (Bing Chat), industry insiders believe that this is likely to help Microsoft arm in arm, in the search market to conquer the city.

Every time a new technology emerges, it’s time for a major industry reshuffle. Although in the global search market in October 2023, Google occupied 91.55% of the market share, while Bing was only 3.11%, which is still a slight decline from last year, but artificial intelligence, a new technological wave hit over, the uncertainty of the game will be greatly increased, which is a huge challenge for the leader Google, while for the follower of Microsoft is a huge opportunity.

Microsoft obviously does not want OpenAI to provide products and services for other customers, the most favourable situation for it is OpenAI only as an enabler behind it, do not come out to the foreground to receive customers. At the same time, Microsoft has other enablers, partners, to ensure flexibility. It’s like a romantic relationship where both parties want to have many lovers and the other party has only one lover of their own.

A Microsoft internal document in March this year shows that in order to persuade customers to choose their own Azure OpenAI service, Microsoft secretly ordered its sales staff to “pull the pedal” OpenAI. when Meta re-released the open source large language model LIama 2, Microsoft announced that it had become the first partner of Llama 2.

Altman is obviously not willing to accept such a “adopted” role, and is also frequently contacting new “gold masters”.

Just when he was confident that he would survive on his own, OpenAI’s board of directors staged a “coup” and drove him out of the company overnight.

II. A for-profit company controlled by a non-profit organisation

While Microsoft has made a large investment in OpenAI, it is investing in OpenAI Global, which was founded in 2019 and is “fully controlled” by the OpenAI non-profit organisation, which was incorporated in Delaware in 2015.

OpenAI Global is “permitted to earn and distribute profits” subject to the mission espoused by the parent organisation, OpenAI, which has always emphasised Microsoft’s “acceptance of our cap-and-trade offer, and the fact that we are keeping the AGI technology and governance retained in the hands of the non-profit and indeed all of humanity”.

In an interview in January 2023, Altman also said, “The future I would like to see is one where access to AI is super-democratised, and there are multiple AGIs in the world that can help people form multiple perspectives without making anyone too powerful.”

Under Microsoft’s investment agreement with OpenAI, Microsoft’s share of the profits gained by OpenAI Global is limited to “gains made prior to the realisation of AGI (Artificial General Intelligence)” as set out in OpenAI’s charter.

“AGI” is defined as “a highly autonomous system capable of outperforming humans in the most economically valuable tasks.” So who decides whether or not to “realise AGI”?

The six members of OpenAI’s board of directors, known as the “Open AI Six”: Chairman and President Greg Brockman, Chief Scientist Ilya Sutskever, Chief Executive Officer Sam Altman, and three independent directors who are not OpenAI employees, Adam D’Angelo, Tyler D’Angelo, and Tyler D’Angelo. Angelo, Tasha McCauley and Helen Toner.

If they believe that AGI has been “realised”, then Microsoft no longer has the right to make a profit. While it is rare for a non-profit board to decide whether shareholders of a for-profit company will receive a profit, OpenAI is unique in that it was founded by Altman and others as a non-profit organisation with a mission to ensure safe general artificial intelligence (GAI) for the benefit of mankind.

They saw the non-profit board as focused on benefiting all of humanity, while the for-profit board served investors. They have also endeavoured to build in another layer of security, which is to keep the board majority independent, i.e., a majority of the members do not hold equity in OpenAI Global.

It is argued that there are indeed some non-profit organisations that own shares in for-profit organisations, most notably the Hershey Trust. But they are in full control of the for-profit companies under them, with no opposition from minority shareholders. In the case of OpenAI, Microsoft’s for-profit interests could directly conflict with the nonprofit interests of the controlling entity.

Once OpenAI achieves its proposed AGI mission, Microsoft fears it will be cut out of the picture. But at last week’s OpenAI Developer Day event, Altman assured Microsoft CEO Satya Nadella: “I highly appreciate the partnership between the two parties on a technical level …… and look forward to working together to achieve the AGI mission.”

In a new interview with Altman in the Financial Times on 13 November, the OpenAI boss said that the company’s partnership with Microsoft is “working well” and that he expects “we’ll be able to raise more money over time.”

III. Idealist Altman

In 2005, Altman dropped out of school with two of his classmates to start a business, developing Loopt, a social media software that enabled them to share their location with their friends, and the famous startup incubator Y Combinator invested in the software.

Three years later, Steve Jobs invited Altman to attend the Apple mobile phone launch event, and the software entered the Apple APP shop. Since then, Altman has developed some more software, all of which were popular, and after operating for a while, sold it. in 2011, he became a partner of Y Combinator.

In the summer of 2015, at a Silicon Valley hotel, Altman hosted a private party to which he invited several industry luminaries, including Tesla’s Musk as well as Musk’s former Paypal partners Peter Thiel and Reid Hoffman.

At the time, they saw that Silicon Valley giant Google had acquired AI company Deep Mind for $400 million, and agreed that if Google and Deep Mind were successful, they would likely have a monopoly on AI technology. They couldn’t let that happen.

On 11 December that year, OpenAI was announced, with funding from Altman, Musk, Hoffman, Thiel and others, with a mission to ensure safe general artificial intelligence (AGI) for the benefit of humanity. At the time of its inception the nature was that of a non-profit organisation, which caught the attention of the industry.

Deep Mind released the sensational Alpha Go the following year, while Open AI was still in its infancy. Altman resigned from Y Combinator in 2019 to become CEO of OpenAI so he could focus on growing OpenAI.

It soon became clear to Altman that achieving the mission would require significant funding, which was hampered by the nature of a not-for-profit organisation, and in March 2019 OpenAI announced the creation of a for-profit entity to ensure that it would be able to raise sufficient funding while retaining the mission, governance and oversight of a not-for-profit organisation. The reason for this “nonprofit + for-profit corporation” structure is that “no existing legal structure that we know of strikes the right balance.”

It is indeed a paradox that in order to achieve his mission of “ensuring safe general artificial intelligence (AGI) for the benefit of mankind”, he has to raise huge sums of money, which needs to be returned, and a non-profit organisation can’t achieve that goal. So another for-profit entity, a limited liability company, had to be formed.

Altman also made an unusual decision:He would not take an equity stake in the company. Having invested in several highly successful tech startups himself, he was already very wealthy and didn’t need the money. Avoiding any equity would help him stay in line with his original mission.

This decision, however, actually turned off some of OpenAI’s potential investors, who suspected that by not taking an equity stake, Altman was not seeing a return on the project. Still, in July 2019, OpenAI received a $1 billion investment from Microsoft.

IV. Is being kicked out of the company you founded a sign, an honour?

As soon as the news of Altman’s dismissal broke, Musk’s social media platform “X” posted a “link to the job application” on its official account, with a “just in case anyone needs it”. Mockery with sympathy, which is very consistent with Musk’s characteristics, perhaps he is throwing an olive branch to Ottoman.

In early 2018, Musk told Altman that he thought OpenAI had fallen a long way behind Google, and offered to take control and run OpenAI himself.Back then, this had happened at Tesla, and Musk took over and was really successful, and he may have wanted to repeat the Tesla experience.

However, this time he didn’t get control, as Altman and the other founders unanimously rejected the offer. So on 20 February 2018, Musk announced that he was stepping down from OpenAI’s board of directors and would no longer be involved in its affairs in any way, citing a conflict of interest between Tesla’s development of Autopilot technology and OpenAI.

At the end of November 2022, ChatGPT exploded in popularity as soon as it was launched, and Musk was furious.On 17 February 2023, he tweeted, “OpenAI was created under the banner of open source, and that’s why I named it ‘Open’ ( ‘Open’) AI, originally a non-profit to be a force against Google, and now it’s become a closed-source, Microsoft-controlled, profit-maximising corporation.”

On 15 March, Musk tweeted again, “I’m still confused as to how a non-profit organisation that I donated about $100 million to in the first place has now become a for-profit company’s with a $30 billion market cap. If this is legit, why isn’t everyone following suit?”

Recently, OpenAI was in talks to sell its current employees’ shares at a valuation of $90 billion, with valuations skyrocketing over the course of the year.

Musk is understandably angry that non-profit organisations suddenly become for-profit companies, and that $100 million, if not a donation to a non-profit organisation, but an investment in a for-profit company, will still get him shares; however, that shouldn’t be the main reason for his anger.

A hundred million dollars is a piece of cake for him, however, he believes that Altman betrayed the original intent, purpose and mission of OpenAI when he helped him found it, which was supposed to be against Google, a potential monopoly in the field of AI, and ended up fuelling a real-life monopoly, Microsoft.

Now that Altman has been kicked out of OpenAI, it proves that Musk was wrong about him and that he and Microsoft are not in cahoots. Instead, it’s just the right time to go to Musk to continue working on AI against the monopoly that is Microsoft.

On 12 July this year, Musk announced in a tweet the establishment of an artificial intelligence company xAI, the company’s goal is to understand the true nature of the universe. 5 November, Musk’s xAI team released its first AI large model product - Grok. according to the introduction, Grok understands the world in real time through the X platform, and is also able to answer questions that are rejected by most of the other AI systems. other AI systems reject tough questions.

Musk will be happy to “take in” Altman, as he himself has twice been kicked out of companies he founded by the board of directors, a traumatic experience that will last a lifetime, and he was kicked out for reasons similar to Altman’s - he wanted to pursue his dream of building an independent, great company, not a third-rate one that was underwritten by the board of directors.

The other reason of course was his pugnacious management style. Either way, he should be easier to empathise with Ultraman.

Musk’s first pot of money came from Zip2, the online yellow pages company he founded. He wanted to buy the domain name “city.com” to compete head-to-head with Yahoo and America Online, but the board of directors preferred to relegate the product to the status of a no-name supplier to newspaper conglomerates. Musk resisted, and the board threw him out, reducing his power at the company and driving him out in disguise.

He was initially prepared to fight to the end, then interim CEO Prodion advised him, “This is your first company, let’s find an acquirer and make some money so you can do a second, third, and fourth company.”

Four years after Musk and his brother Kimball founded Zip2 in January 1999, Compaq bought it for $307 million in cash. 27-year-old Musk received $22 million.

He then started a second company, online payments company X.com, which later merged with Confinity, an online payments company founded by Thiel, Levchin and others, to launch the PayPal service. Till and the others took advantage of his honeymoon to stage a “coup d’état” and kick him out of the company.

Musk felt like he’d been stabbed in the back. “I’m so saddened by this that I can’t even begin to describe it,” he wrote in an email. “I’ve worked my ass off for this company, I’ve got almost all the cash on the books from Zip2, my marriage is on the line, and they say I’m full of shit, and they’re saying I’m full of shit. And they say I’m so full of evil that they don’t even give me a chance to complain.”

PayPal went public in 2002 and was acquired by eBay for $1.5bn in July that year. Musk got about $250 million in return. But it could have grown into a multi-trillion dollar company, and Musk started it back then with a vision of social media + a mega-bank, akin to what we’ve come to call WeChat here. His insistence on acquiring Twitter was to realise this early dream.

He said, “This is the mission that Twitter could be fulfilling in the future. You can create what I think X.com should be if you combine social media with a payment platform.”

Later Musk told Inc. magazine:”Great things are never born in the hands of venture capitalists and professional managers. They don’t have the creativity or the insight.”

Twice in three years, Musk was ousted from the company he founded, a far more harrowing experience than Steve Jobs was ousted from Apple back in the day. Apple deteriorated after Jobs left and finally had to bring Jobs back again to save the day.

I don’t know which script God gave to Ultraman, no matter what, people who have contributed to the development of human technology and progress of civilisation always deserve everyone’s respect, bless him.

In the case of Cheng Yi, for example, from the angel round established in 2018, to the A round in 2020 and the B round in 2023, the only US dollar fund that can be traced in the three rounds of financing is Kang Xi (Delos), and none of the US domestic funds participated.

I. This should be the favourite story of the US dollar fund.

This situation is not general: the founding team of Cheng Yi are all from MNC, and the certainty of what they are doing is also high, such a background, such a story, is the type of project favoured by American PE, VC.

If the sincere benefit can be explained as a team in Shanghai, the United States local funds can not reach (in fact, it is also very far-fetched, even if the domestic biotech, MNC background team to get the United States local institutions financing quite a lot, who do not have a few friends in the United States to do vc docking down).

Then look at another star project Struceture (Shuo Di biological) is obvious: this is also a diet pill star project, background is more prominent, well-known AI pharmaceutical listed company Schrödinger + industry bigwigs Raymond Stevens. Ray itself has rich entrepreneurial experience, or GPCR field well-known scientists, you can go to search the resume.

The company itself is headquartered in the U.S., and it’s still a popular track for diet pills, so it should be able to raise money in the U.S., right? No, Third Rock, MPM, F-Prime, all of them didn’t make any money. Company has been unable to raise money in the United States, it is China’s dollar fund to save it: from its inception in 2016 to the IPO in 2023, the investors are Sequoia China, Stowe Capital, Qiming Venture Capital, WuXi AppTec and so on.

Ditto for checking local US biotech financing news, there are indeed very few diet pill projects. In the U.S. venture capital is highly developed, no company means no one is willing to pay money.

US local institutions are more familiar with overseas new drug market. And the story of diet pills was first validated in Europe and the United States. It can be said that the A-share speculation on the concept of diet pills, than the U.S. stock market is at least 2 years behind. ** Regardless of investment preferences, from the track market space judgement, in the diet drug track, the U.S. fund has more reasons to go out than domestic funds. **

Shodi’s stock price is all the way up after the listing, and I think it must have made the U.S. institutions that didn’t invest very depressed. From 2016 to now, in the domestic institutions have stepped in the weight loss investment opportunities, why the U.S. domestic funds nearly collective missed this track in the weight loss drug?

Is it that U.S. funds do not like to invest in innovative drugs? No, over the past five years the U.S. funds have placed heavy bets in the field of innovative drugs, the scale of which is much larger than the domestic.

ii. Behind it is the difference in the perception of the disease field between China and the United States

Is it because the U.S. fund is not professional enough, can’t understand clinical data, and can’t judge the molecular value? Obviously not, their team is professional enough, and GLP-1 oral small molecule, it is not difficult to judge. There must be a reason behind it:

The point of this, I think, is that the difference in market perceptions of disease areas is really a huge difference between the two major drug markets in the US and China.

In the view of the United States local pharmaceutical investment institutions: the field of oncology is very suitable for biotech survival and development, the field of rare diseases biotech also has a very good opportunity to bend the road to overtake the car, and chronic diseases, especially in the field of metabolism, pharma occupies a dominant advantage, small biotech has no opportunity.

After decades of competition in the U.S. local drug market, there are abundant cases of biotechs having more opportunities: many drugs in the field of oncology come from biotechs, and in the field of rare diseases, there are small companies such as Shire, Alexion, Vertex, and so on, which have grown up to be giants. M&A in these two fields is also the most.

In the view of domestic pharmaceutical investment organisations: the king and his majesty, would rather have seed, big pharmaceutical enterprises do well, I can also go to grab the cake, big pharmaceutical enterprises do not do well, I have to go on.

On the contrary, rare diseases are generally less popular because of affordability issues.

The domestic pharmaceutical industry, so far, has not yet established a pattern, giving biotechs the opportunity to overtake in various disease areas. In the oncology sector, RMB funds are more cautious than USD funds due to the difficulty of setting up a sales force in China.

I think there are 2 reasons why US funds missed the diet drug wave:

• more concerned about oncology projects
• Feel that there are not many opportunities for small companies in the metabolic field

III. The more specialised you are in metabolism, the more scared you are

There is a saying in the investment world that the more you know about metabolism, the more you are afraid to invest in projects in this field. Why? Metabolic field attaches great importance to long-term safety and dosing cycle advantages, phase III requires a large number of people, large investment, and long follow-up time, several MNCs led by Eli Lilly, Novo Nordisk, with the first mover advantage of insulin, have established a dominant hegemony in the field for more than 10 years, which is mainly embodied in the following:

Clinical data crushing: The giants are relentlessly attacking each other with head-to-head clinicals of nearly 10,000 people, showing no mercy, and in this kind of battlefield where billions of dollars are being invested, small biotechs can’t even be counted as cannon fodder. Calculating how rich the current clinical data of Simeoglutide, the amount of money invested behind the scared to death, the small Yue sold itself for a million years may not be enough. biotech financing small hundreds of millions of dollars, completely insufficient ah!

Pipeline layout comprehensive leading: Big pharmaceutical companies on the one hand, the existing drug dosage optimisation, drug delivery cycle extension, improve competitiveness and life cycle, on the other hand, advance the layout of the next generation or even the next generation of drugs, such as the GLP-1 triple target. biotech what to take and MNC to go to the PK, you only have five threes, other people have a straight flush + king of bombs ah.

In the metabolic field, the more specialised you are, the more fearful you are. Especially the dollar funds, most of them know Eli Lilly and Novo Nordisk very well, see the metabolism of new drug projects equal to the bottomless pit of clinical development, the psychology of these two MNC indestructible, naturally unwilling to pay too much attention to metabolism projects.

In a sense, this argument also holds true in the Chinese market, but only for the diabetes field. Koshi also believes that it is difficult to invest in diabetes projects in the domestic market. Because in the diabetes field, also formed a stable pattern, Eli Lilly / Novo Nordisk + Gan & Lee / Tonghua Dongbao, 4 big giants have occupied the market for many years.

Diet drugs are different, due to regulatory approval of diet drugs for indications later than the foreign market, the future data PK will not be as intense as abroad, clinical investment will not be as outrageous as the United States, the Health Insurance Bureau also welcome more entrants, the proportion of domestic fat than abroad, and not as fat as the fat abroad. Under the superposition of all these reasons, the domestic market for weight loss drugs can be said to have just sprouted, and the pattern has not yet been determined, and biotech still has a lot of opportunities. This is also the reason why domestic institutions dare to go out.

**People can only ever earn money within the scope of their own perception, and this, already enough. **When investing in the metabolism field, you still have to be very vigilant. Don’t be like NASH and get directly affected by GLP-1.

In MySQL, the most frequently asked questions are transaction, isolation level, and MVCC, whether it is a large Internet company, a small factory, or even a state-owned enterprise, their coverage rate is as high as 80%.

In fact, the interviewer also knows that everyone will memorize the eight-legged text, but can say that understand, and even say thorough candidates are very rare.

So today I’m going to take you to unlock the black technology hidden in the bottom of MySQL: transactions and isolation.

2. Transactions

2.1 Straight Reward

First, let’s talk about transactions.

Transactions are like a magic show that **ensure that a series of database operations either all execute successfully or none at all. **

Let’s say you’re watching a live stream and you want to reward the hostess with 500 bucks. You need to deduct your account balance and increase the hostess’s account amount at the same time.

If one of the two operations of transferring money fails, then you can lose money or have it disappear and the beauty queen won’t receive it.

This is where transactions come in handy.

It can ensure that both operations either succeed or fail at the same time, there will never be an embarrassing situation of half-success and half-failure.

So, let’s summarize:

• Q: Why do databases have transactions?

• A: In order to ensure that the business runs properly and the data is ultimately consistent.

For those who don’t understand what ultimate consistency is, take a look at this previous post of mine: In-depth: Distributed, CAP, and BASE Theory 3DMzI5Nzk2MDgwNg%3D%3D%26mid%3D2247484896%26idx%3D1%26sn%3D60dd09486fc9ecc652af917d8a311419%26chksm% 3Decac51e9dbdbd8ffc10b79699ea7e4a8fb00aabc743b15cc5c3311970a9e3046592cbb879364%26scene%3D21%23wechat_redirect “http://mp.weixin.qq. com/s?__biz=MzI5Nzk2MDgwNg==&mid=2247484896&idx=1&sn=60dd09486fc9ecc652af917d8a311419&chksm= ecac51e9dbdbd8ffc10b79699ea7e4a8fb00aabc743b15cc5c3311970a9e3046592cbb879364&scene=21#wechat_redirect”)

2.2 Transaction Characterization

Now that you understand what a transaction is and why you need one, let’s talk about the 4 characteristics of transactions.

Let’s talk about the 4 characteristics of transactions: Atomicity, Consistency, Isolation, and Durability, or ACID for short.

Atomicity

Atomicity means that a transaction contains operations that are either all successful or all unsuccessful.

For example, the initial balances of accounts A and B are $800 and $100. At this point, A transfers $500 to B. The breakdown is A account minus $500 and B account plus $500.

The end result is that account A has a balance of $300 and account B has a balance of $600. The operation of updating the balance of these two accounts is either performed in full or not performed at all.

Taking the example of rewarding a beautiful anchorwoman, atomicity ensures that either the money is still there, or the money is transferred to the account of the anchorwoman and the anchorwoman’s thank you brother is rewarded!

Consistency (Consistency)

The state of consistency is maintained before the transaction is executed, and after it is executed.

Two things happen to accounts A and B after a transfer:

1. the money is transferred to account B. At this point, accounts A and B are $300 and $600 respectively;
2. the money is transferred out of the process of database network disconnection, the transaction is rolled back, A, B account or 800, 100 dollars.

In any case, before and after the transaction, the total amount of A and B bank accounts should be $900, which is inconsistent.

Isolation

Isolation is when more than one user accesses the database concurrently, no matter whether it is operating the same library or the same table, the transaction opened by the database for each user can not be interfered by the operation of other transactions, and multiple concurrent transactions should be isolated from each other.

For example, when A transfers money to B, no matter how others transfer money, it will not affect their transactions.

Taking the example of giving a reward to a beautiful anchorwoman, isolation is: no matter how many people are giving the anchor a reward, it will not affect your affairs of transferring money, and it will not affect the anchor to call you a good brother!

Persistence (Durability)

Once a transaction has been submitted, then the change to the data in the database is persistent [i.e., saved to disk] , even in the case of database system failure will not be lost to submit the operation of the transaction.

Take the example of rewarding a beautiful anchorwoman, persistence is: you just transfer money to the anchor, the money into her account, no matter how many voices thank you good brother harvested anchor, the money can not come back.

Next, we **summarize: **

• Q: Why do transactions have these characteristics?

• A: We want to ensure that the data consistency of the transaction, we need some means to achieve, these means are several characteristics of the transaction.

They are atomicity, consistency, isolation, and persistence, where ** consistency is the goal, and atomicity, consistency, and isolation are all means to achieve data consistency**.

3. Transaction Concurrency and Isolation

Transaction concurrency

Concurrency is the ability of a computer system or program to handle multiple tasks or operations at the same time, that is, to allow multiple user processes to work on the same critical area.

For those who want to understand concurrency from a process or processor perspective, see this previous article of mine: [GPM Scheduling Model](https://link.juejin.cn?target=http%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzI5Nzk2MDgwNg% 3D%3D%26mid%3D2247484182%26idx%3D1%26sn%3D6d3f54eea5622a2d7f6323cbb553fdd8%26chksm% 3Decac571fdbdbde09cc8beb982e5df0caafdf5c87587cd3fbd69ca86c33724e9368ab957beac3%26scene%3D21%23wechat_redirect “http://mp.weixin.qq. com/s?__biz=MzI5Nzk2MDgwNg==&mid=2247484182&idx=1&sn=6d3f54eea5622a2d7f6323cbb553fdd8&chksm= ecac571fdbdbde09cc8beb982e5df0caafdf5c87587cd3fbd69ca86c33724e9368ab957beac3&scene=21#wechat_redirect”)

Take the reward anchor for example, concurrency is multiple viewers want to reward the anchor, if you transfer money together, then the anchor account balance how to modify it?

The task here is to transfer money, the user process is the server process responsible for the transaction, and the critical zone is the storage space for the anchor account.

If transaction concurrency occurs, it can cause some unexpected problems, such as the common Dirty Write, Dirty Read, Duplicate Read, and Phantom Read.

Dirty writes

Dirty writing means that during transaction concurrency, one transaction can modify the data of another ongoing transaction, which may result in one write transaction overwriting the data of another write transaction.

When you and Xiao Shuai together to the beauty of the hostess reward, you rewarded 500 dollars, Xiao Shuai rewarded 1,000 dollars, in the write database, you write the data is Xiao Shuai’s data to be overwritten.

The final result is that your money is gone, and the hostess is saying thanks for the reward in the live broadcast!

Transaction isolation

500 bucks is gone, and the hostess is still ignoring you, you’re sad, but you don’t know what to do?

Don’t be sad! Transaction isolation can help you.

MySQL provides transaction isolation levels, including: Read uncommitted, Read committed, Repeatable reads, and Serialization, to solve various concurrency problems in transactions, and to cure all kinds of unhappiness.

RU - Read uncommitted

RU (Read Uncommitted) means that if a transaction starts writing data, another transaction is not allowed to write at the same time, but other transactions are allowed to read this row of data.

RU can exclude writes, but does not exclude read thread implementations.

This isolation level solves the dirty write problem above, but there may be dirty reads, i.e., transaction B reads data that transaction A has not committed.

You want to give the beautiful anchorwoman reward 500 bucks, found that the bank card balance is only 300 bucks, then you thought of a few days ago you borrowed 500 bucks Xiaoshuai, so let Xiaoshuai pay back the money.

Xiao Shuai is very clear about the database isolation mechanism, know that you are in the RU transaction isolation level. So he says he’ll pay you back immediately, and the following scenario occurs:

• Shuai: open transaction A, transfer money to you 500, the transaction is not submitted;

• You: open transaction B, check the balance, and find that the balance has been added 500, so the Shuai Shuai debit note torn off, and ready to give the anchor reward;

• Shuai: see the debit note is gone, so revoke the transaction A. His money is not a penny less, and you only read the balance in his transaction A, but the real balance did not increase, that is, a dirty read has occurred;

• You: The balance of the reward payment is insufficient, and you lose the debit note worth 500 bucks.

You are so disappointed that you plan to cut ties with Handsome and continue to learn the rest of the isolation mechanism to see how to prevent dirty reads from occurring.

RC - Read committed

This level of isolation prevents other transactions from accessing a row of data (both read and write) while it is being written to by one transaction. This ensures that the data read by a transaction is committed, solving the problem of dirty reads.

However, RC will appear ** unrepeatable read ** problem, for example: transaction A need to read the data twice, after reading the first data, there is another transaction B to update the data and submit the transaction.

At this time, transaction A read the data again, the data has changed, that is, ** transaction in the two read data inconsistent **.

In order to salvage his friendship, Handsome transfers you 520 bucks, but he thinks it’s okay to pay you back only 500 bucks, so he asks you to pay him back 20 bucks.

You’ll be too busy watching the show, so you won’t have time to transfer the money. He suggests that you tell him the account password of your bank card, and he’ll only transfer 20 bucks.

To be on the safe side, you open a transaction to check the card balance and tell Shuai the password, the following scenario happens next:

• You: open transaction A, inquired about the bank card balance of 820;
• You: open transaction B, withdraw 800, and submit transaction B. * You: open transaction B, withdraw 800, and submit transaction B;
• You: in transaction A again to check the balance, found that the bank card only 20 dollars, the occurrence of unrepeatable read.

Not only did you not get the money you borrowed, but you lost 280 bucks. The more you think about it, the more angry you get, and you scold the marshal. Then you continue to study the isolation mechanism to see how to prevent the unrepeatable read problem.

RR - Repeatable read

When the same data is read multiple times within the same transaction, no other transaction can access the data (including reads and writes) while the transaction is still open.

This isolation level ** solves the problem of dirty reads and unrepeatable reads ** but there may be ** phantom reads **.

If transaction A reads the data several times, another transaction B inserts or deletes data in the middle of the data rows, then transaction A reads again, you may find that the number of rows of data has changed.

In short, **RR - Repeatable Read ensures that the current transaction will not read other transactions have committed update operations, but can not sense other transactions insert 和 delete operations **.

Shuai knows you won’t lend money again, and was scolded by you, and is resentful. So he thought of using your bank account to mess things up, and the next scenario happened:

• You: open transaction A, want to query the transaction just a few times, the transaction to see the result is 2 times;
• Shuai: open transaction B, found that you can not modify your balance data, so simply to your bank card inside the write 100 times the transaction record, the transaction amount up to tens of millions of dollars, submit transaction B. * You: continue inside transaction A, the transaction amount up to tens of millions of dollars;
• You: inside the transaction A continue to query the number of transactions, found that it became 102 times;

At this point, the police uncle came to the door, said someone reported you malicious money laundering, need to assist in the investigation.

Luckily, after some explanation and investigation through the logs of the bank’s database, it was discovered that someone had maliciously tampered with the transaction records, and you returned home safe and sound.

At this point, you learned the hard way and were shocked to realize that you had made a bad friend! So you sink your teeth into the isolation mechanism.

Serializable

At this isolation level, transactions can only be executed sequentially, solving the problems of dirty reads, unrepeatable reads and phantom reads. However, it is more expensive and has very low performance, and is generally seldom used.

In this case, every time a viewer, like you, wants to give a reward to the anchor, you need to wait in line until the previous transaction transaction is completely finished.

At this point, you learn about the wonders of transactions and the importance of isolation, and plan to learn about databases and stop watching beautiful anchors dance.

Handsome, on the other hand, gets lost further and further down the road of bureau-oriented programming.

4. Summary

Let’s summarize that databases solve the various problems that arise from transaction concurrency through isolation levels:

• RU, read uncommitted solves the problem of dirty writes, but dirty reads may occur;
• RC, read committed solves the dirty read problem, but unrepeatable reads may occur;
• RR, Repeatable solves the problem of unrepeatable reads, but phantom reads may occur;
• Serializable, serialization solves the problem of phantom reads, but performance is low.

How does MySQL implement transaction isolation?

The answer is locking. The higher the transaction level, the more concurrent transactions are solved, which also means more locks are added.

Comparison of number of locks: RU-Read Uncommitted < RC-Read Committed < RR-Repeatable < Serializable-Serialized.

However, frequent locking may result in no way to modify the data when reading it, and no way to read it when modifying it, which greatly reduces the database read and write performance, just like the serialized isolation level.

So, in order to trade off data security and performance, MySQL databases use RR, the Repeatable Read isolation level, by default.

1. Introduction
2. five approaches to solving the massive data problem
3. Classic Scenario Examples
4. Summary

1. Introduction

In recent years, high concurrency, distributed, and big data have become a topic that back-end developers can’t get around. When the recruitment software says that high concurrency, big data, and other project experience is preferred, I’m sure a lot of people are secretly shocked, ** projects are CRUD, and there is no chance to get in touch with these scenarios. **.

However, a great man once said: there are no conditions, to create conditions. Since the work can not be exposed to high concurrency and big data, we can bend the road to overtake the car - usually more attention to similar scenarios in the learning time.

This article describes the common means of solving big data problems, as well as some classic big data scenarios and solutions, which **includes bytes, Ali, Baidu and other major domestic Internet companies interviews with the original questions **. After reading this, I believe that we will be able to be familiar with these big data related problems next time we encounter them on a project or in an interview.

2. Five approaches to solving massive data problems

2.1 Partitioning

It is well known that ** the computational time required for any computer-solvable problem is related to its size. ** The smaller the size of the problem, the easier it is to solve and the less computation time is required. So, for large-scale data problems, we can also divide and conquer and then just merge the results.

Partitioning algorithm is an efficient tool for solving complex problems, it can decompose a problem into several sub-problems, solve these sub-problems one by one, and then combine them together to form an answer to the big problem.

Computer problems dealing with large amounts of data, the idea of partitioning is basically able to solve, but in general will not be the optimal solution, but can be used as a baseline. for example, in some scenarios we may need to gradually optimize the sub-problems to achieve an optimal solution. The traditional subsumption sort is the idea of partitioning, involving a large number of files that can not be loaded into memory, sorting and other problems can be solved using this method.

Classical scenarios of the partition algorithm: large amounts of data can not be loaded into memory, in order to save time and overhead parallel synchronization to solve the problem and so on.

2.2 Hash (Hash)

Hash, hash function, as defined by Wikipedia hash is a method of creating a “fingerprint” of small numbers from any data. In layman’s terms, the hash algorithm allows data elements to be located and queried more quickly.

Hash has a time complexity of O(1) for obtaining a key. Assuming that there are n numbers stored in memory by the hash function, we can obtain the value of a key in a constant amount of time.

With such efficient access to data, hash is clearly a great tool for solving big data problems.

And using hash to solve the data access problem is a pretty brutal way to do it: just put it in and take it out! The only downside to **hash is that it’s memory intensive, and you need to load all the data into memory. **Hash

Hash Scenario: Fast lookups, but need enough memory to hold all the data. In this case, if the amount of data is so large that the memory is not enough to store all the data, it can be combined with partition + hash to deal with it. In fact, we often use this approach when solving real-world problems.

2.3 Bitmap (BitMap)

The core idea of the BitMap algorithm is to use a list composed of bits (bit) bits to record the two states 0-1, and then map the specific data to the specific location of this bit array, bit position 0 means that the data does not exist, set to 1 means that the data exists.

** bitmap through the array to indicate the existence of certain elements, it can be used for fast lookup, reweighting and sorting ** and so on. Bitmap is more widely used in big data scenarios, because it can save a lot of space utilization.

The next two classic scenarios, weighing and sorting to see the effectiveness of bitMap in which it plays a role.

Scenario 1: Finding non-repeating positive integers among 200 million integers

**Bitmap Weighting

Since there is not enough memory to hold these integers, we can use bitmaps. Using a 2-BitMap (each number is assigned 2 bits, 00 for nonexistent, 01 for occurring once, 10 for occurring multiple times), iterate through all the elements, set the position of the element that has been iterated over to 01, and if it is iterated over again set it to 10. Finally, count all the elements that have 01 (occurring only once) in their position.

Scenario 2: Sorting some integer elements [6,4,2,1,5

Bitmap Sorting

Since the elements of the array to be sorted are all less than 8 (and may be much more than that in real scenarios), we can request an 8bit array (one byte) and then iterate over these elements, setting the position subscripts of the existent elements to 1 [0 1 1 1 1 0 1 1 1 0].

After traversing, we again traverse this bit array, you can get the final sorting results:

1
2
3
4
5

for i := 0; i i++ {
   if bitMap[i] == 1 {
      println(i)
  }
}

2.4 Bloom Filter

Bloom filter was proposed by Bloom in 1970, it is actually composed of a very long binary vector and a series of random hash functions.

Bloom filter is mainly used to determine whether the target data exists in a massive dataset and set of intersection. To determine the existence of an example: Bloom filter through the target data mapping, can be O(k) time complexity to determine the existence of the target data, where k for the use of the number of hash functions, which can greatly reduce the traversal of the time required to find.

The workflow of the Bloom filter is shown in the figure above, we next simulate the Bloom filter to do element adding and querying operations.

1. Adding elements
   Bloom filter to add elements, the use of different hash functions on the element value of the hash calculation, so as to get more than one hash value.
   For example, in the figure above, the element 1, after three hashes, after a series of operations to calculate the remainder of the three numbers 3, 5, 6. Then, we will be in the array of 3, 5, 6 three subscripts set to 1, the addition of the element 1 operation is complete.
2. Searching for elements
   In the use of Bloom filters to query the existence of an element, first of all, the given element to perform a number of hash operations, and get the same as when the element is added to the position of the bit array.
   For example, for element 1 to perform three Hash operations, the position of the bit array were 3, 5, 6, to determine whether the value of these positions are all 1, if one of them is 0, the element does not exist; if they are all 1, the probability that the element exists.
3. Element misjudgment
   When the Bloom Filter adds elements, the array position of the element after the Hash may be set to 1 by other elements, for example, after the element x has been Hashed 3 times, the hit position array subscripts are 0, 3, 6, and the values of these three subscripts are set to 1 when inserting the element 1 and element 2.
   Therefore, when querying element x, it will be misjudged that element x already exists, which is the root cause of the misjudgment - multiple hash values of an element may be hit by other elements accidentally, so it is not possible to determine that the element must exist.
   Although BloomFilter has a certain rate of misjudgment, but the probability is very small, about 0.01% ~ 0.05% between, so in some big data scenarios, the use of BloomFilter can significantly improve the efficiency of the judgment, provided that you can accept a small margin of misjudgment.
   2.5 Heap
   Heap is a kind of complete binary tree with two structures: big top heap and small top heap. A big top heap is characterized by its parent nodes being greater than or equal to the values of its left and right child nodes, while a small top heap is the opposite, with its parent nodes being less than or equal to the values of its left and right child nodes.

When building a big top heap or small top heap, we need to use heap sort.

The core idea of heap sort (big top heap example) is: keep comparing the values of all the parent nodes and child nodes, and exchange the larger number into the parent node, after completing one round of comparison, the maximum value of the whole sequence is the root node at the top of the heap.

Then, swap the root node with the end node (the node labeled 8 in the graph), which is then the maximum value.

Then, the remaining n-1 elements are reconstructed into a big top heap, so that the maximum value of the n-1 elements is obtained, and the cycle is repeated to obtain an ascending sequence.

Q: Why don’t we use small top heap for ascending?
A: Because there is no way to guarantee the incrementality of each level. For example, the 5th and 8th elements of the small top heap in the above figure are 26 elements.

In large amount of data, heap sort is a commonly used solution to the TopN problem, and it can satisfy most of the problems of finding the most value.

3. Examples of Classic Scenarios

1) Finding non-repeating integers in a large number of numbers.

Problem Description

Find non-repeating integers out of 10 billion integers. Note that there is not enough memory to hold so many integers.

Question Answer

1. Partitioning + HashMap

10 billion integers, each int integer occupies 4 bytes, 100 x 10^8 x 4B is about 400GB. if we read it into memory as a file, it definitely won’t fit.

So we can split the large file into small files, such as through the Hash operation, will be 10 billion integers assigned to 1000 files, each file to store 10 million integers, these files will be numbered 0 ~ 999, and then they were traversed into the HashMap, where the key is an integer, the value for the number of times it occurs.

Then, the 1000 HashMap were counted out only 1 time in the integer, and then merge the final result.

**Bitmap method

Since integers occupy more memory, we can use one or more bits to mark whether an element occurs and the number of times it occurs, which can greatly save storage space.

We use 2 bits to indicate the state of occurrence of a number: 00 means it has not appeared, 01 means it has appeared once, and 10 means it has appeared many times. 10 billion integers are of int type, each integer occupies 4 bytes, i.e., 32 bits, and the memory required is 2^32 x 2bit = 1GB.

When the available memory is greater than 1GB, the bitmap method can be used to solve this problem. By iterating through the 10 billion numbers, the corresponding subscripts 00->01 (integer occurs once), 01->10 (integer occurs many times), and finally counting the number of 01 (occurs only once).

**Bloom filter

Iterate through the 10 billion integers, they will be deposited into the Bloom filter, before depositing through a number of Hash to determine whether the value of the bit array exists, if it exists that the probability of the existence of the integer; if the value of a Hash does not exist in the bit array, that the integer must not exist.

2) Duplicate URLs found in two large files

Problem Description

Given two files a.txt, b.txt, each holding 5 billion URLs, with each URL occupying 64B, and a memory limit of 4G, find the URLs common to files a and b. Find the URLs that are common to both files.

Answer

The file size of 50 x 10^8 x 64B is about 320GB, so 4G memory is not enough to load all URLs into memory at once.

Partitioning + HashMap

This question is very similar to the previous one, except that the integers are replaced by URLs, and we handle it in a similar way.

First, divide the file where the URLs are stored into multiple small files, so that each small file is no larger than 4G, so that the two small files corresponding to a and b can be processed in memory, and the results can be merged in the end.

First traverse the file a, traversed to the URL by hash mode: hash (URL) % 100, according to the results of the calculation of the URL into a0, a1 … a99.txt, each of which is about 3.2 GB in size. b is traversed in the same way, and the URL is split into b0, b1 … b99.txt. b99.txt.

After splitting, all possible identical URLs are in the corresponding small files, i.e., a0 corresponds to b0, …, and a99 corresponds to b99. All we need to do is to find the URLs that are the same in these 100 pairs of small files.

To find the same URL, we can use HashSet/HashMap, when the URL exists in Set/Map, it means the URL is duplicated, so we can save the duplicated URLs in a separate file, and then merge all the same URLs in the end.

3) Top100 for frequency of request for large session files

Problem Description

There is a file of 100 million conversations, each line of the file is a conversation, the size of each conversation is not more than 1KB, the memory size is limited to 1.5 GB, and it is required to return the 100 conversations with the highest frequency (Top100).

Question Answer

Due to memory constraints, it is not possible to read large files into memory at once. Therefore, the same partitioning strategy is used to split large files into smaller ones for processing.

1. Partitioning + HashMap

First, we use the same way as the previous question to partition the file for processing.

1.5 GB/1 KB is about 1.5 million sentences, and the memory of each file cannot exceed 1.5 million conversations, so we hash each sentence with hash(sentence) % 100, and store the result into a[i] (0<=i<=99) files, each file stores about 1 million conversations.

Then use HashMap to count the 100 most frequent conversations in each file, key is the md5 value of the conversation, value is the frequency of occurrence. The final result is 100*100 = 10,000 conversations, and then the 10,000 words are heap sorted to find out the Top100 frequency conversations.

2. Small Top Heap

Construct a small top heap of size 100, the top element of the heap is the minimum frequency value of the conversation, if in the process of traversing the HashMap, found that the value of the current conversation (the number of times the conversation occurs) is greater than the number of times the top of the heap conversation occurs, then replace the old conversation with a new conversation, and then re-adjust the small top heap.
At the end of the traversal, the 100 words on the small top heap are the 100 conversations with the highest frequency of occurrence of the session we need to count.

4).Extract the highest frequency IP from massive log data

Problem Description

This is an original question from an interview with Baidu, the question is that there is a massive log data containing access IPs saved in a huge file, which can’t be read into the memory, and it is required to extract the IP with the highest number of accesses to Baidu on a certain day.

Question Answer

Partitioning + HashMap

We first use the partition strategy to split the file storing IP data into multiple small files, and then traverse all the small files to extract all the IPs that visited Baidu on that day into a single file.
Then use HashMap to count the number of times each IP accesses, through a variable maxCount to store the value of the highest number of accesses, and constantly compared to find the IP with the highest number of accesses.

5) Count the number of different numbers in a large number of phone numbers

Problem Description

Count the number of phone numbers in a file that contains about 10 billion phone numbers, note that the phone numbers may be duplicates.

Problem Solution

Bitmap method
Since the length of a phone number is 11 digits, and there are 10 cases of numbers on each digit, we need to apply a bit array of length 100 billion (10^11), which requires about 12.5GB of memory 10^11 x 1bit. traverse all the numbers, set the array to 1 when the number occurs, and record the bitmap element that is set to 1 for the first time with a count, and the final result will be obtained at the end of the traversal. After the traversal is finished, the final result will be obtained.

4. Summary

Massive data problems, basically by the above kinds of classical problems morphing, common solution, is the partition, Hash, BitMap, Bloom filter and heap sort. Other types of big data problems may be a combination of various means of solution in different ways.
More complex may be based on text deformation, such as the search engine prefix tree structure, big data text retrieval inverted index and so on, due to space limitations, this paper is not listed.
However, for the classic problems mentioned above, these several solutions are already enough. If you encounter similar interview questions or problem scenarios, we can be handy:)

1. Introduction
2. What are HTTPs?
3. Customizing Certificates
4. Enabling HTTPs on the server side
5. Postscript

1. Introduction

“The global construction team, anywhere, at any time, will put ‘security’ in the first place! And network project security, in turn, is the top priority of Internet products, so can you say a little bit about how data transmission in project development is secured?”

Less than recalling, I only heard the interviewer ask slowly and methodically.

“Heh, construction team, as a new age migrant worker I certainly know that safety comes first. **Wannabe can’t eat fish while constructing! **” At this point, I would like to talk nonsense, but the opposite is a smiling interviewer, not good enough to refuse, so I think for a moment slowly answered: “We are in the project development on-line, in order to interface data transmission security, access may need to use HTTPs”.

At this point there are small partners have questions, data security access, not with POST request can be?

The fact is certainly not the case, “we are in the process of interface design, GET request access security is the worst, because the transfer of data will be spliced in the URL to send directly, while the POST request will be the transfer of data into the Body body, not visible in the address bar, but there is no difference between the two in the transmission process. **To achieve secure data transfer, HTTPs must also be used”. **

2. What are HTTPs?

2.1 Introduction to HTTPs

“So tell me about your understanding of HTTPs”, the interviewer didn’t seem to be very surprised by this answer and asked again.

“In the network, the HTTP protocol communication is ubiquitous, familiar with the RESTful API development program apes know, even if the POST request, in the network transmission due to the reasons such as plaintext transmission, is still not very secure. At this point, we need to use HTTPs, in general terms HTTPs = HTTP + data encryption + authentication + integrity protection.”

The essence of HTTPs is to add a new layer between the HTTP application layer and the TCP transport layer. Assuming that we use the TLS protocol for HTTPs encryption, HTTPs is the communication between HTTP and TLS, and then the communication between TLS and TCP.

2.2 Encryption Process

I went on to say, “HTTPs encryption is roughly divided into two phases, the Certificate Validation and Data Transfer phases, which interact as follows:”

2.2.1 Certificate verification process

1. The browser initiates HTTPs request;

(2) The server generates a pair of public and private keys, the private key is stored by the server itself, and the public key is put in the HTTPs certificate and returned to the client, and the content of the certificate also contains the website address, certificate authority, expiration date, etc. The client verifies the legitimacy of the certificate (certificate validation and data transmission);

(3) The client verifies the legitimacy of the certificate (e.g. whether the URL in the certificate is the same as the current URL and whether the certificate has expired), and if it is not legitimate, it prompts an alarm;

2.2.2 Data Transmission Phase

1. When the certificate is verified to be legal, the client generates a random number as the key of the symmetric algorithm;

2） The client encrypts the random number through the public key and transmits it to the server;

3. After receiving the encrypted random number, the server side decrypts the random number through its own private key;

4. After the server gets the symmetric key (random number), it symmetrically encrypts the returned result data.

3. Homemade certificates

“Okay, so if you were asked to design a server for HTTPs, how would you do it?” The interviewer seemed to want to hear more real-world experience, so he went on to ask.

I was not willing to show weakness, so I added: “After understanding the principle of HTTPs and the authentication process, let’s sort out the three things necessary to open HTTPs secure communication: **CA certificate, server certificate and server private key **.”

1. CA (Certification Authority) certificate is issued by a CA organization to prove the legitimacy of an entity’s identity. 2;
2. the CA certificate needs to be installed on the client machine, i.e. the browser, to verify the authenticity of the server certificate;
3. the server’s certificate is used to send to the client, the server private key to do data encryption and decryption.

“So, when we want to open an HTTPs service, we need to apply for a certificate first. there are two common software for making homemade certificates: OpenSSL and XCA. the former is a command line interface and the latter is a graphical interface. to make sure that the process is visualized, we can use XCA to make the certificate.”

3.1 Creating a certificate with XCA

At this point, the clever and attentive reader may realize: if the interviewer wants to see me applying for a certificate, then there is clearly not enough time! However, in order to let the reader better familiarize with the HTTPs encryption process, the thoughtful program ape ❤ ❤ After the interview, we sorted out the following process:

First, download and install the XCA tool, download address: xca.hohnstaedt.de/ , I downloaded the latest XCA version 2.4.0.

1. After the installation is completed, we open the xca software, click File -> New Database on the main page (select the storage directory of certificate application, and then enter the password)

2. Switch to the certificate page and click on create certificate (I have already created a certificate, so it already exists in the certificate field, so please ignore it)

3. On the source page, select Signature, Algorithm, and Template, respectively, and finally click “Apply all information from the template”

4. switch to the main page, fill in the internal name of the certificate, the rest of the options can be filled in casually, and finally generate a new key

(5) Click Create, OK to return, in the newly generated certificate to create a certificate [Note: When choosing the signature, select “Use this CA certificate for signing”, because we are now applying for server-side certificates, so select the template to choose the default TLS_server].

6）Switch to the main body again, fill in the name and other information, click to generate a new key, OK after the completion of the creation to return to the

At this point, the CA certificate and server certificate are created, and then exported.

3.2 Exporting Certificate

1. On the certificate page, export the CA certificate.

2. Export server-side certificate

3. On the private key page, export the server-side private key

At this point, we are done generating certificates using XCA. Next, turn on HTTPs_ on the server using the certificate and private key.

4. Enabling HTTPs on the server side

The example here is in Golang. Other languages are similar in that you only need to switch to HTTPs when HTTPs are listening for services.

To control whether HTTPs are turned on in Go, you can add three new parameters to the project startup entry:

1
2
3

safeMode := flag.Bool("safe_mode", true, "https mode")
certPath := flag.String("cert_path", "D://runSpace//wecom//ca//server.crt", "server ca")
keyPath := flag.String("key_path", "D://runSpace//wecom//ca//server.pem", "private key")

1）如下图所示

2. safeMode controls whether HTTPs are turned on or not.

The logs are as follows

1

[] Listening and serving HTTPS on :80

3. After the startup is complete, you can type https://127.0.0.1:80 into your browser to access the

Since the CA certificate is not added to the browser, so the access will be prompted, we choose [Advanced -> continue to go] can be.

(4) If you do not want to report this kind of prompt every time you visit, we can add the CA certificate just applied for in the browser, Google Chrome import process is as follows

5）【Manage Certificates -> Import】Import the certificate you just applied in XCA, close the browser and restart.

At this point, the HTTPs service can be accessed normally. If you want to switch to HTTP access, just set the command line parameter safe_mode to false.

The full code is at GitHub.

5. Postscript

The interviewer nodded slightly and asked again, “During the HTTPs encryption process, why are private key authentication and data transmission, two ways of asymmetric encryption and symmetric encryption, respectively, used?”

I smiled wryly, to the interviewer dug a million times the pit, they really all voluntarily stepped up: “This is of course determined by the characteristics of both of them, which:”

• Symmetric encryption: that is, ** both parties to a message use the same key to encrypt and decrypt the message ** using symmetric cryptographic coding techniques. Since the algorithm is public, the key cannot be disclosed to the public. It has a small amount of computation and fast encryption speed. The disadvantages are insecurity and difficulty in key management, such as AES, IDEA.
• Asymmetric encryption: ** can only be encrypted and decrypted by pairs of public and private keys, generally public key encryption, private key decryption **. Process: Party A generates a pair of keys and discloses one of them as a public key; Party B gets the public key, encrypts the data and sends it to Party A, which decrypts it with a dedicated private key. Secure, but slower encryption, such as the RSA algorithm.

“HTTPs encryption uses hybrid encryption, which combines the advantages of symmetric encryption and asymmetric encryption, and transmits the symmetric encrypted public key through asymmetric encryption. In this way, both the security of the public key transmission process [asymmetric encryption] and the efficiency of the data exchange [symmetric encryption] are ensured.”

The interviewer nodded slightly, “Uh-huh, that’s enough of that question for now, next ……” . With a good start, the interview process my legs are not shaking, the mouth is not shivering, **HTTPs also do not want to learn, the evening call on the old silent together to eat fish! ** emmm really good day ah ~~

2. swagger introduction

2.1 Background

Years later, facing the code bouncing on the screen, the back-end developer Xiao_❤️ will recall the distant night when he first wrote the interface document. On that day, little ❤️ had just finished writing the code, looking at the evening sun in the sky, thinking that everything was fine and he could get off work soon. Front-end developers small A came over, “hello old brother, may I ask how is the interface development, tomorrow can be linked up?”

“Emmm although this piece of work is more difficult, but I have been working day and night these days to catch up with the work, tomorrow should be able to interconnect!”

As soon as Little A heard this, his eyes began to shine, thinking that he could get off work earlier once the intermodulation was over in a couple days. So he started to say excitedly, “Give me an interface document quickly, I’ll add them to the page tonight”.

At this time, the development manager also came over, “interface documents or very important, involving the consistency and stability of the front and back end interfaces, the need to write and maintain, small ❤️ son you add off work, write the document tonight, it does not take long, estimated half an hour to get it done! In the process of writing to pay attention to the systematization of thinking, think about what you do this thing where the value point? What you do, and the company’s differentiation of other teams in the where? What you do, does it precipitate a set of reusable physical information and methodology? Why are you the one doing it and no one else can? Sink your thinking into a daily, weekly, and monthly report that reflects thinking about the interface documentation thing, not just progress. Okay, get busy!”

The leader has spoken, how can we say no? Even though 10,000 grass horses are running in my heart, I can only nod my head with a smile and say yes.

As a result, each unimpressive interface needs to be defined by in-parameters, out-parameters, and examples, so that it can be linked with the front-end and subsequently archived. After much deliberation over the format of the documentation, a final version was created. The documentation for a particular interface looks like this:

Interface description: add xxxx business.
Interface address: /api/v1/resourcexx.
Request method: POST

_Request header parameters _
Field name Type Required Description Content-Type string yes application/json;charset=UTF-8

Request body parameters: _
Field Name Type Required Description app_id string Yes application_id source_type integer Yes data_source, 0: local, 1: data-center data_ids string No data_set_id _No

…………

_Response parameters: _
field name type description code integer response code, 200 for return success, 400 pass parameter error, 500 internal error msg string _error message, when the response code is 200, the value is “ok” _ data json _response body, according to the different business to decide the _

Request example:

1
2
3
4
5
6
7
8
9

curl --location --request POST 'http://127.0.0.1:5187/nlp/addMarketBoard' \
--header 'Content-Type: application/json' \
--form 'file=@".../测试小组挂断率.xlsx"' \
--form 'file_name="002"' \
--form 'component_type="0"' \
--form 'app_id="1056"' \
--form 'board_name="minio测试"' \
--form 'source_type="0"' \
--form 'data_ids="67"'

Some people can not help but ask: this is something that should not take too much time, right ~ Indeed, as our leader said, an interface does not look too much text, are some fixed format, copy and paste the work only.

However, the end of development in addition to my own, there are dozens of interfaces developed together with colleagues ah …… The leader’s idea is that I’ll do it together with the things that I do by hand anyway.

There is no CPU I’m not too sure, but when the last interface added to the document, I saw the bright moon in the night sky, and then fell into meditation: I do not know whether it was to stay in love with the beauty of the evening sun, or pondering Wang Jiefu “when the moon shines on me back” state of mind.

Anyway, I don’t want to do this job of writing interface documents for the second time in my life. So, before it was 11 o’clock (lights out at 11 o’clock in the company), I started to search for tools that can automatically generate interface documents, and I came across it - swagger.

2.2 What is swagger?

Swagger is the development tool framework for the world’s largest OpenAPI specification (OAS), and one of the world’s most popular restful API documentation generation tools, its advantages are:

• support for cross-platform , cross-language
• community open source, and very active
• has a very complete ecosystem (Swagger Editor, Swagger Codegen, Swagger UI …)

restful API: a kind of industry common routing specification design, it is the style of the Internet all the data are regarded as resources, the request URL naming is to locate the resource, the specific operation of the resource by the request method [GET/POST/DELETE/PUT] to decide. For example:
GET http://localhost:8080/api/v1/resource // Get a certain resource on a certain local server.

2.3 Introduction to the swagger tool

This time, using Go as an example, we use swaggo as the tool to automatically generate the swagger API documentation and gin-swagger as the rendering package implementation of the Swagger UI.

The three toolkits to be introduced are as follows:

go get -u github.com/swaggo/swag/cmd/swag or go install github.com/swaggo/swag/cmd/swag@latest (after go1.70)
go get github.com/swaggo/gin-swagger
go get github.com/swaggo/gin-swagger/swaggerFiles

More on when to use them next.

3. Building the Swagger Sample Project

3.1 Build a new project

Operating environment:

1. Windows 10 (OS optional)
2. Goland
3. Git
4. Go 1.17.11 (optional)

The whole project operation is realized in Goland, we default readers have installed the Goland/Go SDK, and downloaded the Git tool to execute the command operation.

3.1.1 Creating a New Project and Introducing gin-swagger

First, let’s create a new project in Goland and name it swagger-test:

Then, go ahead and introduce the go mod into the project to manage the dependencies we’ll be downloading later:

go mod init swagger
go mod tidy

Next, introduce the gin-swagger middleware and the built-in files for swagger file management into the project

go get github.com/swaggo/gin-swagger
go get github.com/swaggo/gin-swagger/swaggerFiles

3.1.2 Writing Server Listening Code and Comments

Since gin-swagger has been introduced for this purpose, we will use the gin framework to handle browser requests in the following.

First, create a new api package and add the Hello method and annotations, which are the interface methods we will use later:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

package api

import (
    "net/http"

    "github.com/gin-gonic/gin"
)

type Response struct {
    Code    uint32      `json:"code"`
    Message string      `json:"message"`
    Data    interface{} `json:"data"`
}

func Hello(c *gin.Context) {
    res := Response{Code: 1001, Message: "first interface, hello", Data: "connect success!"}
    c.JSON(http.StatusOK, res)
}

Among them, the interface annotations in lines 11~25 are manually added according to the definition of swagger, the swaggo tool automatically generates API documentation, these annotations will be used, the basic definition of the annotations are:

• @Summary, interface summary
• @Description, the interface description
• @Tags, the interface tags, used to group the API
• @Accept, the interface accepts the type of input parameter, support mpfd (form), json, etc.
• @Produce, the type of output parameter returned by the interface, supports mpfd (form), json, etc. * @Param, the type of input parameter.
• @Param, the definition of the input parameter, from front to back are:

As shown in the code @Param user_id query string true “User ID” minlength(1) maxlength(100), @Param format is:

1. parameter name 2. parameter type 3. data type 4. whether field is required 5. parameter description 6. other attributes

Path and response comments about the interface have:

• @Success, specify the data of the success response, in the format of 1.HTTP response code 2.response parameter type 3.response data type 4.other description

• @Failure, specifies the data after a failure response, same as Success

• @Router, specifying the route and HTTP method.

More fields can be found in the Swaggo documentation: [[gitcode.net/mirrors/swa…](https:// gitcode.net/mirrors/swaggo/swag/-/blob/master/README_zh-cn.md)]

Next, create a new main.go in the project’s home directory, specifying the api/v1/hello interface to point to the Hello method in the api package:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

package main

import (
    "swagger/api"

    "github.com/gin-gonic/gin"
    swaggerFiles "github.com/swaggo/files"
    ginSwagger "github.com/swaggo/gin-swagger"
)

func main() {
    r := gin.Default()
    r.GET("/swagger/*any", ginSwagger.WrapHandler(swaggerFiles.Handler))

    v1 := r.Group("/api/v1")
    {
       v1.GET("/hello", api.Hello)
   }
    r.Run(":8080")
}

Since we are going to generate the swagger UI page through the main method, we need to add comments to the page definition above the main method, common fields are:

• title, the title of the swagger UI.
• version, the version of the interface
• description, the description of the swagger document
• host, the host address of the record
• BasePath, the base path, which will be displayed in the swagger UI with host automatically spliced in.

After defining the swagger UI page with the above comments, we need to specify a route to access it:

1
2

r := gin.Default()
r.GET("/swagger/*any", ginSwagger.WrapHandler(swaggerFiles.Handler))

gin.Default() defines an initialization engine for gin, and /swagger/*any serves as the access route. After generating the swagger file and starting the service, you can access the swagger UI page in a browser via 127.0.0.1:8080/swagger/index.html, with /swagger being the path prefix.

Next, we’ll add a new business access interface. We’ll add a new route group /api/vi as a prefix, and under this route group we’ll add a /hello interface, which will point to the Hello method under the api package. Finally, we specify an 8080 port to listen on.

3.2 Generating API Documentation with the swaggo Tool

3.2.1 Installing the swaggo tool

Once you have defined all of this, you can install the swaggo tool to automatically generate the swagger files. There are many ways to do this, but we’ll use the Git method here.

We need to download and install git on our computer first, here we have git installed by default and configure the environment, then install swaggo:

(2.0) Linux installation:

go get -u github.com/swaggo/swag/cmd/swag (before go1.70)
go install github.com/swaggo/swag/cmd/swag@latest (after go1.70)

If the installation is unsuccessful in Linux environment, it may be caused by the inconsistency between GOPATH path and GOROOT, you can use vi /etc/profile command to modify it:

export GOROOT=/usr/local/go ##GoLang installation directory.
export PATH=$GOROOT/bin:$PATH
export GOPATH=/usr/local/go ##GoLang project directory

Then refresh GOPATH: source /etc/profile

2.1）windows Installation method (this method is selected by default below)：

go get -u github.com/swaggo/swag/cmd/swag

2.2）macOS Mounting method:

mv $GOPATH/bin/swag /usr/local/go/bin

Test if the installation was successful:

swag -v

3.2.2 Generating API Documentation with swaggo

After the installation of swag is complete, we go to the home directory of the swagger-test project [that is, the root directory where the project was created, in the example, D:\runSpace\swagger-test], and use the command to generate the swagger file:

swag init

At this point, we can see that the project has automatically generated a docs directory with the newly generated swagger files:

./docs
├── docs.go
├── swagger.json
└── swagger.yaml

3.2.3 Importing API Documentation from the docs Package

Next, we add the path to the just-generated docs package in the main.go file, and import this path so that the API documentation can be accessed after the service is started:

Then, we start the project with go run main.go:

You can see that port 8080 is up.

4. Swagger Document Rendering Demonstration and Testing

4.1 Swagger UI Page Access

4.1.1 Server Connectivity Testing

First, we access the hello interface in the browser:

http://127.0.0.1:8080/api/v1/hello

Successful access indicates that the port is up.

4.1.2 Swagger UI Page Access

Next, we open the swagger UI page to view the API documentation for the interface.

Path：http://127.0.0.1:8080/swagger/index.html

As you can see, the swagger UI page is now accessible. It contains a basic description of swagger and API interface information, and you can also view the doc.json interface definition online:

Here, we can see the power of swagger, just through a few comments and route definitions, you can write what we need to define in the interface document.

Moreover, the swagger API documentation is universal, almost all front-end, back-end and testing staff can understand. If there is a modification of the interface in the development process, we only need to change the comments, and then use the swag init command to update the docs, and then restart the service. Then we can dump the updated docs http://ip+port/swagger/index.html to the front-end or testers.

4.2 API Testing

In addition to API documentation, swagger UI also provides interface testing, we can see a Try it out button on the right side of the /hello interface “Parameters” on the homepage:

点击之后可以输入测试用例，输入参数后点击 Execute 执行：

The execution steps are:

1. Click Try it out button to start the test. 2.
2. Input the interface input parameter, there is only one input parameter in the example, and it is not business related, so you can fill in the parameter as you like.
3. Click the Execute button to execute the test.
4. Swagger UI generates a curl command to call the server-side interface. 5.
5. Return the response result: success/fail.

See here, you must have understood the basic usage of swaggo tool. However, the API documentation is generally only needed when we do testing in the development environment, after the service is online, most of the interface documentation is to be invisible to the user. Moreover, updating the swagger documentation every time a service goes live will also affect the speed of our service compilation.

Therefore, we can use the flag library’s command parsing to control whether or not to hide the swagger UI interfaces at runtime.

5. The flag library controls whether to render the Swagger UI.

5.1 Adding flag parameter control

To control whether or not to render the UI page, we use a bool type field called swagger. The code for the main function above reads:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

func main() {
    swaggerTag := flag.Bool("swagger", false, "Whether to generate swagger document at build time")
    flag.Parse()

    r := gin.Default()
    if swaggerTag != nil && *swaggerTag {
       r.GET("/swagger/*any", ginSwagger.WrapHandler(swaggerFiles.Handler))
   }

    v1 := r.Group("/api/v1")
    {
       v1.GET("/hello", api.Hello)
   }
    r.Run(":8080")
}

Using the flag library’s command parsing, you can pass in swagger commands to control whether or not to expose the swagger UI interface each time you run the service.

First, compile the main function into an executable called main.exe, and then run the main.exe executable (do the following in git):

go build main.go
./main.exe

The service started successfully.

5.2 Accessing a flag-controlled server

5.2.1 Without adding swagger parameters

Next, we access the Swagger UI page by entering the address http://ip+port/swagger/index.html into our browser. We find that the server returns the response code 404: Interface does not exist.

It seems that the swagger UI page cannot be accessed without specifying the swagger parameter at runtime.

5.2.2 Adding the swagger parameter

Next, let’s add the swagger parameter at runtime to turn on access control for the UI page:

. /main.exe -swagger=true

As you can see, the swagger UI access was successful.

6. Summary

Summary of Swagger auto-generated docs:

• When using the swag init command to generate documentation, the docs directory must be referenced in the service main, otherwise the API information cannot be accessed after running;
• Swagger UI access is controlled by ginSwagger.WrapHandler, which can be turned on or off by using the flag library parameter command;
• After modifying the interface definition, you need to use swag init to update the swagger file immediately, and then restart the service to take effect.

Swagger automatically generates documentation from the interface annotations, remember a few commonly used can be used in conjunction with the official documentation to use the best. This idea of annotative programming makes our code more readable, and can be seen in widely used frameworks in the industry, such as Spring Boot, Mybatis, and so on.

At the end of the day, development is a mental job, but more than that, it’s a physical job. Therefore, the lazy programming, automated programming approach I personally respect. Extra time to brush the blog, look at the technical articles do not smell it ~ and then do not help, go to the forum to see the black talk discourse, the future to fool the leadership or CPU interns may be useful (🐕)

3. https://link.juejin.cn?target=https%3A%2F%2Fblog.csdn.net%2Fcsdnno11%2Farticle%2Fdetails%2F119634795 “https://blog.csdn.net/csdnno11/article/details/119634795“)

1. Common Sorting Algorithms
2. Getting Started
3. Sorting Chained Tables
4. Merge intervals
5. Advanced Topics:
6. kth largest element in an array
7. Finding the median of two positively ordered arrays
8. Summary

1. Common Sorting Algorithms

Common sorting algorithms include Insertion Sort [O(n^2)], Summation Sort [O(nlogn)], Heap Sort [O(nlogn)], and Quick Sort [O(nlogn) -> O(n^2)].

If you are not familiar with the sorting algorithms, you can learn them on Visualgo Algorithmic Data Structures, which has videos and algorithm definitions, so I won’t repeat them in the article.

2. Introductory topics

1) Sorted Chained Lists

Power button topic 148:

This is leetcode question 148, which is simple: sort an unordered linked table, with the advanced requirement: sort the linked table in O(nlogn) time complexity and constant-level space complexity.

Among the sorting algorithms The algorithms with O(nlongn) time complexity are subsumption sort, heap sort, and fast sort, but the worst time complexity of fast sort is O(n^2) which is not applicable in this problem, and heap sort is a better sorting algorithm for chained lists compared to subsumption sort.

So, we use the subsumption sort to realize this question, which is based on the idea of Partition Recursion with the following steps:

1. divide the chain table from the middle node into two, you can use fast and slow pointer way to realize;
2. sort the two sub-chained tables separately;
3. the sorted list for the merger, you can get a complete sorted list.

The Go code is as follows:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38

func sortList(head *ListNode) *ListNode {
    if head == nil || head.Next == nil {
        return head
    }
    // 快慢指针，找到链表中间节点
    slow, fast := head, head
    for fast.Next != nil && fast.Next.Next != nil {
        fast = fast.Next.Next
        slow = slow.Next
    }
    next := slow.Next
    slow.Next = nil
    return mergeTwoList(sortList(head), sortList(next))
}
​
// mergeTwoList 合并两个子链表
func mergeTwoList(h1, h2 *ListNode) *ListNode {
    pre := new(ListNode)
    hair := pre
    for h1 != nil && h2 != nil {
        if h1.Val < h2.Val {
            pre.Next = h1
            h1 = h1.Next
        } else {
            pre.Next = h2
            h2 = h2.Next
        }
        pre = pre.Next
    }
    if h1 != nil {
        pre.Next = h1
    }
    if h2 != nil {
        pre.Next = h2
    }
    return hair.Next
}

The code is very simple and efficient to write in recursive fashion:

2) Merging intervals

Force Buckle Topic 56:

The general idea is to merge an array of intervals so that there are no overlapping intervals after the merge. Suppose there are intervals [1,3] and [2,6], which should be merged into [1,6].

The idea of solving this problem is to compare the 2nd element of all the subarrays with the 1st element of the array that follows, and if there is a crossover, find a larger element to be the 2nd element of the new array. That is, when comparing [1, 3] 和 [2, 6], we first compare 3 is greater than 2, so we can see that there is a crossover between the two arrays, and 6 is greater than 3, so the 2nd element of the new array is 6 => [1, 6].

How can we be sure that the first element is unique? For example, if we have 2 arrays: [3, 5], [1, 8], we’ll have to determine the size of the first element when we merge them, so for simplicity we can sort all the arrays first, making sure that the array with the smaller first element comes first. In Go, you can implement array sorting with simple functions:

1
2
3

sort.Slice(arr, func(i, j int) bool {
    return arr[i][0] < arr[j][0]
})

Next, we can write the full code:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

import "sort"
​
func merge(intervals [][]int) [][]int {

    sort.Slice(intervals, func(i, j int)bool{
        return intervals[i][0] < intervals[j][0]
    })
    var res [][]int
    for _, interval := range intervals{
        var size = len(res)

        if size > 0 && interval[0] -1][1] {
            res[size-1][1] = max(res[size-1][1], interval[1])
        }else{
            res = append(res, interval)
        }
    }
    return res
}
​
func max(x int, y int) int {
    if x < y {
        return y
    }
    return x
}

More than 94% efficient golang commits:

3. Advanced Topics

1) Finding the Kth largest element of an array

Power Buckle topic 215:

K largest element of the array, this question is the Internet factory interview questions, I have been tested in the byte, Baidu interview, so you must be proficient.

The meaning of the topic is very simple, is to find the kth largest element of the array sorted, such as [1,2,3,4,5], k = 2, then the second largest element is 4. If there is a topic with the same elements, we do not have to pay attention to, for example, [5,5,4,3,1], k = 2, then the second largest element is 5, rather than 4.

The advanced requirement is: design and implement an algorithm with O(n) time complexity to solve this problem.

This problem is a classic one in sorting algorithms, in which, if we don’t require time complexity, we can just use all kinds of sorting algorithms to first sort the array and then get the kth largest element. However, whether in interviews or in daily use, we need to pursue the optimal time complexity of the algorithm, and efficient sorting algorithms are heap sort (nlogn), subsumption sort (nlogn) and quick sort (nlogn -> n^2).

How to satisfy the time complexity O(n)?

Quick sort

The answer is mentioned in 9.2 of Introduction to Algorithms, Third Edition, where the randomized_select algorithm is a selection algorithm that expects linear time, and is modeled after fast sort, where the idea is still recursive partitioning. Unlike fast sort, however, the randomized_select algorithm deals only with one side of the partition, which we describe in code:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

func findKthLargest(nums []int, k int) int {
    rand.Seed(time.Now().UnixNano())
    return quickSelect(nums, 0, len(nums)-1, len(nums)-k)
}
​
func quickSelect(a []int, l, r, index int) int {
    q := partition(a, l, r)
    if q == index {
        return a[q]
    } else if q < index {
        // 根据partition元素，只处理划分的一边
        return quickSelect(a, q + 1, r, index)
    }
    return quickSelect(a, l, q - 1, index)
}
​
func partition(a []int, l, r int) int {
    // 选取随机数作为边界元素
    i := rand.Intn(r-l+1) + l
    a[i], a[r] = a[r], a[i]

    x := a[r]
    i := l - 1
    for j := l; j < r; j++ {
        if a[j] i++
            a[i], a[j] = a[j], a[i]
        }
    }
    a[i+1], a[r] = a[r], a[i+1]
    return i + 1
}

Submission of results:

We found that the optimized algorithm randomized_select time complexity for fast sort exceeds 99% of users.

Heap Sort

Next we implement it again with heap sort, which is a sorting algorithm designed to take advantage of the data structure heap. A heap is a structure that approximates a complete binary tree and simultaneously satisfies the heap form: i.e., a child node’s key value is always less than (or greater than) the key value of its parent node:

• Small top heap: each node’s value is less than or equal to the value of its child node;
• big top heap: each node’s value is greater than or equal to the value of its child node.

This question is to find the largest kth number, so we build a big top heap, each time we find the largest number, remove the largest number (implementation: the top of the heap and the end of the heap elements are interchanged, and then the size of the heap will be subtracted by one), and then find the k-1th number, the code is as follows:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

func findKthLargest(nums []int, k int) int {
    heapSize := len(nums)
    buildMaxHeap(nums, heapSize)
    for i := len(nums) - 1; i >= len(nums)+1-k; i-- {
        nums[0], nums[i] = nums[i], nums[0]
        heapSize--
        maxHeapify(nums, 0, heapSize)
    }
    return nums[0]
}
​
func buildMaxHeap(nums []int, heapSize int) {
    for i := heapSize / 2; i >= 0; i-- {
        maxHeapify(nums, i, heapSize)
    }
}
​
func maxHeapify(nums []int, i, headSize int) {
    l, r, largest := i*2+1, i*2+2, i
    if l < headSize && nums[l] > nums[largest] {
        largest = l
    }
    if r < headSize && nums[r] > nums[largest] {
        largest = r
    }
    if largest != i {
        nums[i], nums[largest] = nums[largest], nums[i]
        maxHeapify(nums, largest, headSize)
    }
}

As you can see, heap sort has a time complexity of O(nlogn), which is a bit more time-consuming than randomized_select for fast sort:

2) Find the median of two positively ordered arrays

Force Buckle topic 4:

The general idea of the question is to find the median of two ordered arrays nums1 and nums2, and requires a time complexity of , which assumes that the two arrays will not be empty at the same time.

For this problem, the first and easiest way to think of is to merge the two arrays and take out the median. However, the operation of merging the arrays has a complexity of m+n, which does not meet the requirements of the question. Seeing the time complexity of log, we associate it with binary search. 1.

1. When the total number of two arrays is odd, the middle number is the median; when the total number is even, the mean of the two middle numbers is the median. 2;
2. know the length of the array, to find the median of the two arrays combined, only need to bisect the search to find the middle position of the array of 1 or 2 elements;
3. As the location of the middle number to determine, so the two arrays only need to ** bisect the search for an array ** can be, in order to minimize the complexity of our search for the length of the smaller arrays;
4. the most critical: bisection cut should be how to determine whether the end, we assume that there is an array A and array B (length of m and n, respectively), the first bisection cut on the array A, that cut point i, and j [due to i + j = (m + n)/2, so j = (m + n)/2 - i] need to be satisfied, * * A [i], B [j] the left-hand side of the elements should be smaller than the right-hand side of the elements, which can be guaranteed that i and j are one of the median elements**.

The graphical representation is as follows:

1 two-point split:

2 two-point splits:

After splitting again, left = 3, right = 3, and the dichotomous cut of array A is complete. At this point, the cut subscripts for arrays A and B are i = left = 3, j = mid - left = 2, respectively.

The median result will only come from these i, j, i-1, j-1 numbers. 1. when the total number of arrays is odd, the median is the larger of the left partition (i.e. i-1 and j-1):

1. when the total number of arrays is odd, the median is the larger of the left partition (i.e., i-1 and j-1);
2. when the total number of arrays is even, the median is the average of the largest number in the left partition and the smallest number in the right partition.

In addition, we have to consider boundary cases, such as when i, j is 0, the subscript i-1 or j-1 does not exist; when i, j is m, n, the subscript i or j does not exist. Since the title states that m and n will not be 0 at the same time, at least one of i and j must exist.

The Go code is as follows:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52

func findMedianSortedArrays(nums1, nums2 []int) float64 {

   m, n := len(nums1), len(nums2)
   if m > n {
      return findMedianSortedArrays(nums2, nums1)
   }
   left, right, mid := 0, m, (m+n+1)/2
   for left < right {
      i := (left + right + 1) / 2
      j := mid - i

      if nums1[i-1] > nums2[j] {
         right = i - 1
      } else {
         left = i
      }
   }
   i, j := left, mid-left
   var leftMax float64
   if i > 0 && j > 0 {
      leftMax = float64(max(nums1[i-1], nums2[j-1]))
   } else if i > 0 {
      leftMax = float64(nums1[i-1])
   } else {
      leftMax = float64(nums2[j-1])
   }
   if (m+n)%2 == 1 {
      return leftMax
   }
   var rightMin float64
   if i < m && j < n {
      rightMin = float64(min(nums1[i], nums2[j]))
   } else if i < m {
      rightMin = float64(nums1[i])
   } else {
      rightMin = float64(nums2[j])
   }
   return (rightMin+leftMax)/2
}
​
func max(a, b int) int {
    if a>b {
        return a
    }
    return b
}
​
func min(a, b int) int {
    if areturn a
    }
    return b
}

Submission of results:

4. Summary

A few of the classic questions in the Power Buckle Sorting category are above, and their probability of being encountered in interviews or written exams is very high. In Huawei’s machine test questions, the frequency is also very high, except for the last question, which is a bit less difficult, the rest of the questions basically appear at least once in 10 interviews.

1. Introduction
2. RDB
3. AOF
4. Comparison of Two Persistence Mechanisms
5. Summary

1. Introduction

Q: What is Redis?

A: Redis, or Remote Dictionary Server, is an in-memory cache database written in C and widely used in Internet products.

Whether domestic or foreign, from the top 500 companies to small startups are using Redis, many cloud service providers also Redis-based caching services, message queuing services, and memory storage services, when you use these services, in fact, is to use Redis.

When you use these services, you’re actually using Redis.** As a developer, you’re bound to be asked about it in interviews, even if you don’t use it on the job! **

_Q: Why is Redis so important and what are some common application scenarios for it? _

A: The reason lies in the fact that Redis is a purely in-memory operation with high execution efficiency and supports rich data structures itself. It has a wide range of application scenarios, including but not limited to ** caching, event publish or subscribe, distributed locking ** and so on.

_Q: Are all Redis operations in-memory? _

A: No. The single-threadedness of redis means that it is a single-threaded operation when receiving client IO requests for reads and writes. But there are multi-threaded scenarios for redis itself, such as asynchronous deletion, persistence, and cluster synchronization.

Redis as a common knowledge point in the interview, the common eight-legged text must have been familiar with. However, in today’s increasingly voluminous Internet market, relevant knowledge points, can we answer the depth and breadth of the interviewer wants?

For example, today we want to review the knowledge point, Redis persistence mechanism.

_Q: What are the common Redis persistence mechanisms? _

A: RDB and AOF.

2. RDB

2.1 Introduction

RDB (Redis Database Backup file), or snapshot mode, is the default data persistence method in Redis.

RDB is actually a timer event within Redis that checks the times and times of changes to the current data every once in a while to see if they meet the trigger conditions specified in the configuration file.

When the conditions are met, Redis creates a child process through the operating system call fork(), which ** shares the same address space, file system, and semaphores** as the parent process. After that, Redis traverses the entire memory space through the child process, copies the data set to a temporary file, and when the copy is complete, notifies the parent process to replace the original file with the new RDB file to complete the data persistence operation.

At the same time, during the persistence process, the parent process can still provide services to the outside world, and the parent and child processes realize the separation of data segments through the multi-process COW (copy and write) mechanism of the operating system, so as to ensure that the parent and child processes do not affect each other.

2.2 Summary of advantages and disadvantages

During RDB persistence, the Redis fock child process saves all Redis data to a newly created dump.rdb file, which is a resource-consuming and time-wasting operation. This is a resource-consuming and time-consuming operation. Therefore, Redis servers should not create rdb files too often, or the performance of the server will be seriously affected.

In addition to this, the biggest shortcoming of RDB persistence is: There can be a lot of data loss during the last persistence process. We imagine a scenario, in the process of RDB persistence, Redis server suddenly down, then the child process may have generated rdb file, but the parent process has not yet had time to use it to cover the old rdb file, the buffer in the file has not been saved, will lead to a large number of data loss.

The advantage of RDB data persistence is that it restores very quickly, so it is more suitable for large-scale data recovery. If you are not particularly sensitive to the integrity of the data (allowing for the loss of data during the persistence process), then RDB persistence is very suitable.

3. AOF

3.1 Introduction

AOF, append only log file, is also known as append mode, or log mode.

AOF logs all write commands executed by the server, and only those commands that modify memory, and Redis writes these commands to the appendonly.aof file at regular intervals.

We can re-execute the AOF file to restore the dataset when the server starts up, a process known as command replay.

3.2 Three Persistence Mechanisms

When Redis receives a modification command from a client, it will first perform the corresponding checksum, and if the command is error-free, it will immediately store the command in a buffer, and then append the buffer data to the .aof file at a certain rate.

In this way, even if there is an unexpected downtime, you only need to store the command in the aof file and perform a “command reenactment” to restore to the state before the downtime.

In the above execution process, there is a very important part of the command write, which is a disk IO operation: in order to improve the write efficiency, Redis does not write the content directly to disk, but first puts it into a memory buffer, and then only when the buffer is full or meets the persistence policy of the AOF does it actually write the content in the buffer to the disk (fsync operation). disk (fsync operation).

There are three persistence strategies (i.e., the frequency of fsync operations) for AOF:

• always: every time the server writes a command, it calls the fsync function once to write the command inside the buffer to the hard disk. In this mode, a server failure will not result in the loss of any command data that has been successfully executed, but its execution speed is very slow;
• everysec (default): the server calls the fsync function once every second to write the commands in the buffer to the hard disk. In this mode, if the server fails, the command data executed within one second at most will be lost, and it is usually used as the AOF configuration policy;
• no: the server does not call the fsync function, the operating system decides when to write the commands in the buffer to the hard disk. In this mode, when the server encounters unexpected downtime, the number of lost commands is uncertain, so this strategy, uncertainty is greater, and is not commonly used.

Redis is still at risk of losing data if the data in the cache is not written to disk before experiencing downtime. The number of commands lost depends on when the commands are written to disk: ** The earlier the commands are written to disk, the less data will be lost in the event of an accident. **

Since is fsync is a disk IO operation, it is slow! If Redis has to fsync once (ALWAYS) to execute a command, it will severely impact Redis performance.

In production servers, Redis usually fsyncs every 1s or so by default (everysec) to maintain high performance and minimize data loss.

The last strategy (no), which lets the operating system decide when to synchronize data to disk, has many uncertainties and is not recommended.

Note: The sync and fsync functions are two functions provided by the operating system to prevent data inconsistencies in caches and files caused by “delayed writes”.
The sync function puts the modified data into the cache write queue and returns without waiting for the end of the IO operation.
fsync, on the other hand, waits for the end of the IO operation before returning, and ensures that modified blocks are written to disk immediately to ensure that the file data is consistent with the cache.
i.e., the Linux fsync() function flushes the contents of a given file from the kernel cache to the hard disk synchronously, and sync() operates asynchronously.

3.3 Rewrite Mechanism

Under the AOF persistence policy, Redis runs for a long time, and the aof file gets longer and longer. If the machine is down and restarted, the command to reenact the entire aof file will be very time-consuming, which will cause Redis to be unable to provide services to the public for a long time.

Therefore, in order to keep the size of the aof file within a reasonable range, Redis provides an AOF rewrite mechanism, that is, the aof file for “thinning”: the Redis server can create a new AOF file to replace the existing AOF file, ** the old and the new files save the same database state, the difference is that the new file does not have the task redundancy commands ** so the size of the file will be smaller than the old file, the old file will not have the task redundancy commands ** so the new file will be smaller than the old file. The difference is that the new file does not have task redundancy commands**, so the file size is much smaller than the old one.

Redis provides two ways to rewrite an AOF file: manually execute the BGREWRITEAOF command, or configure a policy to rewrite it automatically. AOF file rewriting is similar to the RDB persistence process, which involves forking a child process to manipulate the original AOF file.

As shown in the figure: the parent process continues to process new requests during the AOF rewrite process. If new commands are added, they are appended to the AOF rewrite buffer and then directly appended to the new AOF file.

Comparison of AOF and RDB

RDB persistence mechanism AOF persistence mechanism full backup, save the whole database incremental backup at a time, save only one command to modify the database at a time each time to execute the persistence operation of the interval is longer, the interval of saving is one second by default (everysec) data is saved in a binary format, its restore speed is faster using text format to restore data, so the data restore speed is generally execute the SAVE command will block. SAVE command will block the server, but manually or automatically triggered BGSAVE will not block the server AOF persistence will not block the server at any time.

Before Redis 4.0, we could only choose RDB or AOF as the persistence mechanism; after Redis 4.0, we can configure Redis persistence to be a hybrid mechanism, i.e., RDB+AOF are both used as persistence methods, for details, please refer to the redis.conf file:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

rdbcompression yes

rdbchecksum yes

dbfilename dump.rdb

appendfilename "appendonly.aof"

appendfsync everysec

auto-aof-rewrite-percentage 100
auto-aof-rewrite-min-size 64mb

aof-rewrite-incremental-fsync yes

rdb-save-incremental-fsync yes

If for data recovery, there is both an RDB file and an AOF file, we should recover the data through the AOF file first, as this maximizes the security of the data.

5. Summary

2023 has come and gone, and the days of the Internet expanding like it did in previous years and grabbing a share of the application windfall are gradually passing! If the tide of the computer industry is receding for a short or long period of time, what will determine whether programmers will “dry swim” on the beach? Is it marginal business, or is it an aging crisis?

I do not think so, in fact, the rising tide and the ebbing tide is a trend, some people catch the crisis early, early to wear a good bathing suit to the shore; some people are not enough to the status quo, to the sea deeper swim. They are the wise man in the Internet wave, the wise man will not worry about the wave receding, because they have been prepared in advance. And the opportunity is always in favor of these early prepared people!

Young people, let’s join hands and dogpile under the wave of the Internet together~

On the weekend, I drove to the beach with my friends. Those who have been to Yangmeikeng should know that you need to take a speedboat to go there from Yangmeikeng to Lukzui Mountain Resort.

Worthy of Shenzhen tour attractions on the top 5 places, 4 or 5 pm when the queue for the boat is still very many people, buy a good ticket we were called to a shore steps waiting to get on the boat, the scene is slightly chaotic.

**The crowds were a bit heavy, but there weren’t a lot of boats arriving to take passengers. **

Just as I was generally sweating for the staff maintaining order, I saw them go back and forth ordering several groups of people to get those people on the boat in an orderly fashion.

Not long after that, a thin, dark, middle-aged man came to call us, saying that a boat could hold 10 people, so he ordered the 10 people in front of us to stay put, and the 10 people who had been ordered could get on the boat.

Sure enough, software design comes from life, and this is the classic data consumption problem in system design!

2. Message middleware

When the amount of data (passengers) is too much, the system (the speedboat carrying passengers) can not consume the data immediately, it will put the data into a consumption queue (the shore ladder) and wait for it, playing the role of a traffic peak shaving.

** In distributed systems, a major way to implement consumption queues is to use message middleware **.

What is Message Middleware

Message Broker is an infrastructure component for delivering messages, notifications and events in distributed systems.

It allows data and information to be exchanged asynchronously between different components, applications, or systems to enable clipped, decoupled, and scalable communication.

The fundamentals of message middleware include the following key concepts:

1. Message Producer (Producer): This is the sender of the message, usually an application or component, which sends the message to the message middleware.
2. Message Consumer: This is the receiver of the message, usually an application or component, which receives and processes the message from the message middleware.
3. **Message Queue: ** This is the core component of the message middleware, it is a queue to store messages, message producers will be put into the queue, message consumers from the queue to get the message. The message queue usually uses the first-in-first-out (FIFO) principle.
4. Message Topic (Topic): In addition to message queues, message middleware also supports message topics, which allow publish-subscribe model of message communication. Message publishers publish messages to topics, while subscribers can subscribe to specific topics to receive related messages.

Advantages of message middleware include:

• Decoupling: Message middleware allows producers and consumers to operate independently; they do not need to be directly aware of each other’s existence. This decoupling makes the system more flexible and maintainable.
• Scalability: By increasing the capacity of the message middleware, it can easily handle more message traffic and consumers.
• ** Asynchronous Communication:** The message middleware allows asynchronous communication, where producers can continue to work without having to wait for a message to be processed, thus improving the performance and responsiveness of the system.
• Message Persistence: Messages are usually persisted so that they are not lost even if the message middleware or consumer fails.

There are many different implementations and protocols for message middleware, and some of the popular ones include ActiveMQ, RocketMQ, RabbitMQ, Kafka, and others.

They have different characteristics and advantages in different usage scenarios and requirements.

Message middleware is widely used in a variety of applications, including ** microservice architecture, big data processing, real-time data analysis, log collection, event-driven architecture ** and so on.

Next, we introduce common messaging middleware and their advantages, disadvantages and applicable scenarios to help you make an informed choice in application development.

3. ActiveMQ

Features:

• ActiveMQ is a Java-based open source messaging middleware that implements the JMS (Java Message Service) specification.
• Support for multiple messaging models , including peer-to-peer and publish-subscribe .
• Provides high availability and load balancing , supports master-slave replication , can be used to build high-availability systems .
• Suitable for Java applications, but there is some support for clients in other programming languages.

Benefits:

• Easy to use for rapid development and prototyping.
• Integrates with the Spring framework for easy integration with Spring applications.
• Suitable for small to medium sized systems and intra-enterprise communications.

Disadvantages: * Relatively low performance.

• Relatively low performance, not suitable for high throughput and latency demanding scenarios.
• Does not support large-scale message flow, not suitable for big data and real-time analysis applications.

** Applicable Scenarios:** ActiveMQ is suitable for internal communications that require simple messaging and small to medium-sized systems. It performs well in intra-enterprise communication and lightweight applications, but is not suitable for high performance, high throughput and large-scale data processing.

In general, ActiveMQ domestic Internet companies landing less , mostly traditional enterprises in use.

4. RocketMQ

Features: ** RocketMQ is an open source MQ framework from Alibaba.

• RocketMQ is Alibaba’s earlier open source MQ framework , based on the Java language written , and later donated to Apache , is a fast , reliable , scalable distributed messaging middleware .
• Supports publish-subscribe and peer-to-peer messaging models.
• With high performance and low latency , suitable for large-scale messaging .
• Supports a rich set of client-side languages, including Java, C++, Python, Go, and so on.

** Advantages:** * High performance and low latency, suitable for large-scale messaging.

• High performance and low latency, suitable for high-throughput large-scale applications.
• Supports multiple messaging models, suitable for different business scenarios.
• With powerful monitoring and management tools.

Disadvantages:

• Deployment and configuration are relatively complex and require some specialized knowledge.
• The community is relatively small , compared with some other messaging middleware , documentation and ecosystem is relatively immature .

** Applicable Scenarios:** RocketMQ is suitable for large-scale applications that require high performance, low latency, and scalability, such as e-commerce platforms, financial systems, and Internet of Things applications.

5. RabbitMQ

Features:

• RabbitMQ is an open source messaging middleware that implements the AMQP (Advanced Message Queuing Protocol) specification.
• Supports a wide range of messaging models, including peer-to-peer, publish-subscribe, and RPC.
• Provides reliability messaging , support for transactions and message acknowledgment .
• Has multiple client libraries and supports multiple programming languages.

Benefits:

• Mature technology with high stability, widely used in enterprise applications.
• Provides high availability and load balancing mechanisms.
• Supports multiple programming languages, suitable for cross-language applications.

Disadvantages:

• Relatively low performance, not suitable for large-scale applications with high throughput.
• Deployment and configuration complex, need some learning costs.
• itself is erlang language development, source code is more difficult to analyze, need solid erlang language skills.

** Applicable Scenarios: ** RabbitMQ is suitable for enterprise-level applications that require reliability and transaction support, but do not have particularly high performance requirements.

6. Kafka

Features:

• Kafka is a high-throughput, low-latency distributed messaging middleware for large-scale data processing and real-time stream processing.
• Mainly used in the publish-subscribe model to store messages as logs.
• Highly scalable and available , suitable for building large-scale real-time data streaming applications .
• Support for a variety of clients, including Java, Python, Go and so on. ** Pros:** ** Cons:** ** Applicable Scenarios:** Kafka is suitable for applications that require high throughput, low latency, and large-scale data processing, such as log collection, real-time data analytics, and event-driven architecture.
  • Complex to deploy and configure, requires specialized knowledge.
  • Not suitable for small-scale applications, high relative complexity.
  • High throughput and low latency for large-scale data processing and real-time stream processing.
  • High scalability and support for building large-scale data pipelines.
  • Data persistence and data replication to ensure data reliability.

7. Technology Selection

RabbitMQ and Kafka

RabbitMQ and Kafka are two of the most commonly used messaging middleware, the main differences between them are:

• Performance: The performance of message middleware is mainly measured by throughput. Kafka’s stand-alone QPS can reach the million level, RabbitMQ’s stand-alone QPS is in the 10,000 level, and kafka is even higher;
• Data Reliability: kafka and rabbitMQ are equipped with multi-copy mechanism, data reliability is higher;
• Consumption mode: Kafka is actively pulled by the client, and RabbitMQ supports two modes: active pull and server push. So RabbitMQ message real-time is higher, and for the consumer is more simple; and kafka can be pulled by the consumer according to their own situation, throughput is higher;
• ** idempotency:** kafka supports idempotency for a single producer, single partition, single session, while RabbitMQ does not support it;
• Other features: RabbitMQ supports priority queues, delayed queues, dead message queues (store queues of messages that can’t be consumed), and more.

How to choose the right messaging middleware

In application development, choosing the right messaging middleware depends on specific requirements:

• If your application is a small to medium-sized system with low performance requirements, and is more concerned with ease of use and rapid development, then ActiveMQ may be a good choice.
• If you need to handle large-scale messaging and are looking for high performance and low latency, then RocketMQ or Kafka may be a better fit, depending on your application type and requirements.
• If your application is an enterprise application that requires reliability and transactional support, but does not require high performance, then RabbitMQ may be a good choice.
• The final choice also depends on your technology stack, the experience of your team, and specific business requirements. It is recommended that you carefully evaluate your application requirements before choosing a messaging middleware and make your choice on a case-by-case basis.

Of course, no matter which messaging middleware you choose, you need to have an in-depth understanding of its features and how to use it to ensure that it can meet the application requirements to build an efficient and reliable distributed system.

8. Conclusion

Regardless of which message middleware is used, we can often see the wonderful use of consumption queues in our daily lives.

**With these buffering methods, our daily travel and consumption order can be well protected. **

1. Machine learning
2. Labeling systems
3. Number of system tasks
4. Flow of system tasks
5. Summary

1. Machine Learning

1.1 What is machine learning?

As the name suggests, the ability and insights of a computer program or system to learn by itself without direct human help or intervention, so that it can answer questions for humans. In layman’s terms, it is the process of creating an Artificial Intelligence (or AI for short) model that learns on its own like a human being, allowing for rapid advancement of knowledge in a particular area.

The difference is that the AI only needs a short period of self-learning to reach or even exceed the human level. For example, AlphaGo only learned Go for 9 days before it could beat a Korean 9-dan player with a 99% win rate, so it is undoubtedly a god of learning.

1.2 Artificial Intelligence is not intelligent

Although AI’s learning ability is very strong, in the eyes of human beings, it is at the level of a god of learning.

What’s interesting is that most outsiders, including me, have been unaware for a long time of the fact that AI itself is not intelligent, e.g., machine learning, which often relies on low-paid crowdsourced workers to manually annotate and fine-tune the data so that they can continue to learn.

And I happen to be working on big model-related data labeling lately, so I’ll share with you some of the labeling-related points I’ve recently combed through.

1.3 What is Data Labeling

Data annotation is the process of attributing, labeling, or classifying data to help machine learning algorithms understand and categorize the information they need to process. This process is critical in model training for AI, allowing AI to accurately understand various data types such as images, audio files, video clips, or text.

The current decision-making AI (e.g., Shake or Taobao’s recommendation system, Tesla’s Smart Driving, etc.), generative AI (e.g., OpenAI’s ChatGPT, Baidu’s Wenxin Yiyin, etc.) can’t be separated from the annotation step, and all of them currently rely on a large number of annotators, or are screened by human beings after being annotated by machines.

Here to share a text annotation GitHub annotation project.

2. The annotation system

This annotation system is called Open Assistant, a project that aims to make chat-based language macromodeling accessible to everyone.

This project has been open sourced on GitHub, ⭐️ has broken 32k in just a few months, for those interested: [github.com/LAION-AI/Op…]

On this annotation system, several processes of machine learning can be simulated, including conversation tree expansion, session annotation and final scoring, and then filtering out the conversations with the highest scores to continue learning.

2.1 Conversation tree

In the annotation system, the conversation tree is the most basic data structure, which simulates the conversation process of machine learning, and its structure is as follows:

First, the root node of the tree is the Initial Instruction, which is the first sentence given by a user (Prompt), which may be a prompt, or may throw out a question.

Second, then Chatbot gives a reply, chatbot based on different processing dimensions, including their own level of knowledge, summarization ability, etc., will give a reply to the user’s question (Assistant Response).

Third, when the chatbot gives a reply, ** the user continues to ask ** (Prompt Response), at this time the dialog tree has come and gone.

Then the chatbot gives a response (Assistant Response), continuing to expand the dialog tree.

When a node in the tree, or the user, no longer responds, the tree ends. However, these chatbots will continue to learn from all the generated dialog trees, i.e., all the conversations with all the users, in order to become smarter and give better answers during the conversations.

2.2 Tasks

In addition to the dialog tree, Task is also a very important data structure in the annotation system. The update of the dialog tree and its nodes is accomplished through one different task. Tasks in the annotation system are of the following types:

• Initialization prompt (initial_prompt) task: the system requires the user’s identity to create an initial instruction, this initial instruction may be to seek an explanation of the concept, a problem of algebraic operations, or the creation of the article type of requirements, that is, ** the user asked a question; **
  label_initial_prompt: let the labeler to label the initial instruction, when more than one person labeling is completed, start by the chatbot to reply to the initial prompt, that is, ** let the labeler to screen out unhealthy questions, such as illegal and criminal, politically sensitive questions; ** ** the chatbot reply to the initial prompt, that is, ** the user asked a question; ** the user asked a question; ** the user asked a question; ** the user asked a question; ** the user asked a question; ** the user asked a question
  ** Chatbot reply (assistant_reply): replying to the initial instruction with a trained chatbot can be seen as the first question answer, at the second level of the dialog tree. ** This step is used by different AIs to reply to the user’s question according to their knowledge level; **
• Labeling the chatbot’s reply (label_assistant_reply): the content of the chatbot’s reply to the user’s instructions/replies is labeled, and the labeled qualified replies are ready to proceed to the next round of the conversation or to the sorting task (enough people are labeled qualified). ** labeling personnel to screen and score the AI’s responses, and ultimately select the AI with the best responses to continue training (can be interpreted as raising compulsion, eliminating unqualified ones); **
• Reply as a user (prompter_reply): reply to the chatbot’s message as a user, corresponding to the dialog tree will add a node, that is, ** by the user to continue to ask questions **;
• Label the user’s reply (label_prompter_reply): label the content of the user’s reply to the chatbot, and the next round of dialog can be carried out after the labeling passes, ** label the user’s questions, and screen out unhealthy and meaningless information **;
  ** Sorting users’ replies (rank_prompter_replies): when many users have replied, each user’s annotation result is qualified, at this time ** the annotator sorts the text quality of users’ replies and selects some optimal replies for the bot to learn; **
• Ranking the replies of chatbots (rank_assistant_replies): similarly, through ** sorting scoring, the optimal replies among chatbots are selected for the models of other bots to proceed to learning and training; **
  ** Random tasks: the system randomly selects more than one task for the annotators to complete.

3. The number of system tasks

The state flow of the conversation tree and its node expansion is accomplished by one different task. Since the number of session trees in different states is inconsistent, the number of different types of tasks displayed on the page is also inconsistent.

1) Number of random tasks (random)

The number of random tasks (shown on the page as “I feel lucky”, this is a system bug, don’t worry) is the total number of all task types, when you choose to do a random task, the system will randomly select a session tree and execute the corresponding task.

2) Number of tasks to write commands as a user (initial_prompt)

In order to minimize write conflicts when updating data tables, the system will limit the number of active conversation trees, and the number of conversation trees that can currently remain active is the number of tasks to write instructions as a user.

This number is controlled by a combination of two indicators:

• Metric 1 is the number of dialog trees remaining to be created, plus the number of dialog trees that have already been created and can continue to grow;
• Metric 2 is the number of dialog trees currently available for queuing;

Finally, the number of tasks writing instructions as a user takes the smaller of the two metrics.

** In simple terms, this number is the number of questions the user can and does ask. **

3) Number of tasks labeled user initial instructions (label_initial_prompt)

The dialog tree needs to be labeled with the user’s initial instructions after they have been written by the user. Therefore, the number of tasks for labeling the user’s initial instructions is the number of dialog trees that are currently active, have a status of initial waiting to be labeled (initial_prompt_review), and have a user’s number of labels that is lower than the configured value.

When multiple users have completed labeling a dialog tree and the labeling results are satisfactory. The state of the dialog tree changes to growing, and the number of tasks labeled with the user’s initial instruction is reduced by one.

** In a nutshell, this number is the total number of initial questions that can be annotated by the annotator. ** ***

4) Number of tasks to reply as a chatbot (assistant_reply)

When 3) has finished annotating the user’s initial instructions and needs to reply to the session tree, it will first reply as an assistant. Therefore, the number of tasks replied to as a chatbot is the number of conversation trees that are currently active, growing, and created with the role of prompter.

**This is the total number of labeled questions that the AI can reply to (of course, in the system, a human user can also simulate the AI). ** ** This is the total number of questions that the AI can reply to.

5) Number of tasks labeled for chatbot replies (label_assistant_reply)

Once 4) replies to the conversation tree as a chatbot, these replies need to be labeled. Therefore, the number of tasks to label chatbot replies is the number of conversation trees that are currently active, with a conversation tree state that is growing (GROWING) and below the configured value. These conversation tree leaf nodes are created with the role of chatbot (assistant).

** In a nutshell, this quantity is the total number of annotated AI replies that are available to the annotator. **

6) Number of tasks replying as a user (prompter_reply)

After 5) has finished labeling the chatbot’s replies, if the labeling result is qualified, it needs to continue to reply to the session tree, and here it will reply as a user.

Therefore, the number of tasks replying as a user is the number of active, growing conversation trees whose leaf nodes were created with the role of chatbot (assistant).

7) Number of tasks to label the user’s replies (label_prompter_reply)

After 6) replies to the conversation tree as a user, these replies need to continue to be labeled.

Therefore, the number of tasks labeled user replies is the number of conversation trees that are currently active, with a conversation tree state that is growing (growing) and below the configured value, and for which the creation role of the leaf node of the conversation tree is the user (prompter).

8) Number of tasks to rank chatbot/user replies (rank_assistant_replies/rank_prompter_replies)

When the annotation is completed, if the number of nodes (both root and leaf nodes) in the conversation tree reaches the target value, the status of the conversation tree is set to ranking.

Therefore, the number of tasks to be ranked is the number of conversation trees that are currently active and have the status of being ranked. There are two types of ranking tasks:

1. if the leaf node of the dialog tree is a user reply, then the user’s reply is ranked;
2. sorting the replies of the chatbot if the leaf node of the dialog tree is a chatbot reply.

4. Labeling system task flow

1) Writing commands as a user

Before writing a command as a user, some basic checks are done, such as user authentication, whether the number of queued tasks for the user in the recent period exceeds the limit value, and whether the remaining number of creation of the dialog tree is sufficient.

If one of the basic checks fails, the process is terminated; otherwise, the initial instruction writing starts.

When the user completes the task of writing the initial instruction, the details of the initial instruction (e.g., text, user name, role-prompter, time, etc.) are stored in the database as a new dialog message and a new dialog tree is generated, which is now in the state of initialization prompt (initial_prompt_review).

2) Labeling initial user commands

After the dialog tree is created, the user’s initial instructions need to be labeled.

Before labeling, it is necessary to do some basic checks, such as user authentication, whether the number of queued tasks of the user in the recent period exceeds the limit, and whether the number of remaining tasks for labeling the user’s initial instruction is sufficient.

If one of the basic checks does not pass, then report an error and return; otherwise, start labeling the user’s initial instructions.

Then from inside the conversation tree randomly select a conversation tree that needs to be labeled with initial instructions (prompts_need_review), and label the sentences of the initial prompts for scoring.

The dimensions for labeling are commonly:

• Whether it is spam;
• Not Chinese, inappropriate, contains PII, hate speech, or contains sexual content;

Scoring dimensions are:

• high quality, level of creativity, level of humor, level of politeness, level of violence

When the annotation is submitted, the system will save the annotation results of each user, and store the annotation results in different tables according to the specifics of the annotation [e.g., whether to report or not, whether to like or not, etc.].

Next, it will determine whether the number of times the dialog tree has labeled the user’s initial instruction reaches the configured value, and if a certain number of labeling times has been reached, it will start to rate the sub-node of initial instruction.

Note: the current scoring is only based on whether it is spam or not, and the language is not clear (not Chinese) to calculate the score.

Score calculation method: suppose the initial instruction of the dialog tree has been labeled by 3 users, 1 of them labeled it as spam, and the other 2 users labeled it as not spam, the ratio of labeling as spam is 1/3. Similarly, if 1 user labeled the initial instruction as language-unintelligible, and the other 2 users didn’t label it as language-unintelligible, the ratio of labeling as language-unintelligible is 1/3. The total score of labeling is 0.34 (1/1), the total score of labeling is 0.34 (1/1), the total score of labeling is 0.34 (1/1). The total marking score is 0.34 (1-1/3-1/3).

In the end, when the labeling score exceeds the configured value (default 0.6), the labeling result of the root node is set as qualified and stored in the table to start the next step, otherwise, it is set as unqualified, and the dialog tree is set to the state of aborted_low_grade, and will not be continued in the subsequent process.

3) Replying as a chatbot

After the initial command is labeled, the dialog tree turns into a growing state, at which time the dialog tree can be replied by the chatbot.

Before replying to the initial command as a chatbot, some basic checks are done, such as user authentication, whether the number of queued tasks of the user in the recent period exceeds the limit, and whether the remaining number of replies in the dialog tree is sufficient.

If one of the basic checks fails, the process is terminated; otherwise, a conversation tree that has been labeled initial instruction completion and is in a growing state is randomly selected for replying.

When the chatbot replies to a task, the details of the reply (e.g., text, username, role-assistant, time, etc.) are stored in the database table as a new dialog message, and the current chatbot’s reply task is updated to done.

4) Labeling chatbot replies

After the chatbot replies to the initial command, it needs to annotate the reply.

Before labeling, some basic checks should be done, such as user authentication, whether the number of queued tasks for the user in the recent period exceeds the limit, and whether the number of remaining tasks for labeling the chatbot’s replies is sufficient.

If one of the basic checks is not passed, an error is reported and returned; otherwise, from inside the conversation tree randomly select a conversation tree that needs to be marked for chatbot replies (reply_need_review), and start marking chatbot replies.

In addition to the same marking dimensions as the marking initial instruction, the marking dimensions for marking chatbot replies are increased:

• As a response to the prompt task, is it a bad response?

Scoring dimensions added:

• How helpful was it?

When the annotation is submitted, the system saves the annotation results for each user and stores the annotation results in a different table according to the specifics of the annotation [e.g., whether it is reported or not, whether it is liked or not, etc.].

Next, it determines whether the number of times the conversation tree has labeled the user’s initial instruction has reached the configured value, and if a certain number of labeling times has been reached, it starts to rate the sub-node of bot replies.

Note: Like labeling the initial instruction, the current scoring is only based on whether it is spam or not, and the language is not clear (not Chinese).

Eventually, when the labeling score exceeds the configured value (default 0.6), the labeling result of the root node is set as qualified and stored in the table to start the next step, otherwise, it is set as unqualified, and the conversation tree is set as aborted_low_grade, and will not continue to go through the subsequent process.

When the labeling is completed, determine whether the number of nodes in the current conversation tree has reached a certain number, and if the target value is reached, the state of the conversation tree will be set to the sorting (ranking) state.

5) Replying as a user

After the bot replies have been labeled, the dialog tree continues to grow, and then replies as user are used to expand the child nodes.

Before expanding the tree as a user, some basic checks are done, such as user authentication, whether the number of queued tasks of the user in the recent period exceeds the limit, and whether the remaining number of replies in the tree is sufficient.

If one of the basic checks fails, the process is terminated; otherwise, a bot replying to a conversation tree that has been labeled complete and is in a growing state will be randomly selected for replying.

When the chatbot replies to a task, it stores the details of the reply (e.g., text, username, role-assistant, time, etc.) as a new dialog message in the database table and updates the current user’s reply task as done.

6) Labeling user replies

After the user has replied to the dialog tree, it is necessary to continue to label the user’s replies.

Before labeling, it is also necessary to do some basic checks, such as user authentication, whether the number of queued tasks of the user in the recent period exceeds the limit, and whether the number of remaining tasks for labeling the user’s replies is sufficient, etc. If one of the basic checks is not passed, an error is reported.

If one of the basic checks does not pass, an error is reported; otherwise, from the conversation tree randomly select a conversation tree that needs to be labeled with the user’s reply (reply_need_review), and start labeling the user’s reply text.

As with the labeling initial command, the marking dimensions for labeling user replies are:

• Whether it is spam or not;
• is not Chinese, inappropriate, contains PII, hate speech, or contains sexual content;

The scoring dimensions are:

• quality, creativity, humor, politeness, violence.

When the annotation is submitted, the system will save the annotation results of each user, and according to the specifics of the annotation [e.g., whether to report or not, whether to like, etc.] the annotation results will be stored in different tables.

Next, determine whether the number of times the dialog tree annotated user replies reaches the configured value, and if a certain number of annotations has been reached, start scoring the sub-node of user replies.

Note: Like the annotation initialization command, the current scoring only calculates the score based on whether it is spam or not, and the language is not clear (not Chinese).

Eventually, when the labeling score exceeds the configured value (default 0.6), the labeling result of the root node is set to qualified and stored in the table to start the next step, otherwise it is set to unqualified, and the conversation tree is set to aborted_low_grade, and no longer continue to go through the subsequent process.

When the labeling is completed, determine whether the number of nodes in the current conversation tree has reached a certain number, if it reaches the target value, then the state of the conversation tree will be set to the ranking state; otherwise, the conversation tree will continue to grow, and the cycle of steps 3) to 6).

7) Ranking chatbot replies

After completing the task of 4) labeling the chatbot’s replies, we need to sort the chatbot’s replies if the conversation tree enters the sorting state.

Before sorting, some basic checks will be done, such as user authentication, whether the number of queued tasks for the user in the recent period exceeds the limit value, and whether the number of remaining tasks for sorting chatbot replies is sufficient.

If one of the basic checks fails, an error is reported and returned; otherwise, a conversation tree waiting for chatbot replies to sort (incomplete_rankings) is randomly selected from inside the conversation tree to start sorting tasks.

Sorting involves ranking the correctness and validity of multiple replies from the chatbot.

When the sorting results are submitted, the system saves the sorting results for each user and records the sorting operation in the log table by adding 1 to the number of sorted messages for the sorted message node.

When the number of sorted chatbot replies for a conversation tree node exceeds the configured value, scoring ranking starts for the sorting of that conversation tree, and the state of the conversation tree at this time is the ready_for_scoring state (ready_for_scoring).

After sorting by the algorithm, the ranking of chatbot replies of the sub-nodes of the dialog tree is calculated, and if the scoring process is completed successfully, the dialog tree is changed to the ready_for_export state; otherwise, if there is an error in the scoring process, the dialog tree will be set to the scoring_failed state.

5. Summary

In addition to text annotation, we also come across image, audio, and video annotation during AI model training, which is done by a large number of annotation employees. However, as AI and GPT grow well, and large models bloom both at home and abroad, the labeling may be left to robots in the future, and the labeling will slowly become smarter and more automated.

According to market research, the current application of AI in the entertainment media field is dominated by content distribution, and there are some auxiliary applications in the content production stage, and in the later stage, it will move towards massive auxiliary content creation or even massive replacement of human creation.

• Machine-assisted human stage: generative AI greatly reduces the cost and threshold of content production, reduces costs and increases efficiency for content companies, and existing large model companies are expected to obtain higher profits;
• Machine “replacement” stage: users only need to input instructions to get the required content created by AI, the importance of the content distribution link decreases, the existing Internet entertainment giants face the challenge of “accurately providing content that meets user needs” to “providing content production tools that meet user needs”.

To put it in human terms, now these AIs are tools for us to keep our rice bowls, but in the future, they may be tools for us to steal our rice bowls!

Hello, everyone, thank you for coming, I am the main character today - Go!

The Birth of Go

In September, 2007, in Google Labs, 61-year-old Rob Pike, 74-year-old Ken Thompson, and 74-year-old Rob Pike were working on the Go language. Rob Pike (Rob Pike), 61, Ken Thompson (Ken Thompson), 74, and 64-year-old Rob Thompson (Ken Thompson) in Google Labs. Ken Thompson, 74, and Robert Griesemer, 64. Ken Thompson, 74, and Robert Griesemer, 64, three engineers who got together and gave me the simple name of Go.

As they were both parents and had a son at an older age, I was always “loved by all”. And because all three of them are very powerful, there is Ken, the father of B, C, and Unix, who won the Turing Award and the U.S. National Technology Award, Rob, the creator of UTF-8 character encoding, and the main contributor to Google V8 and HotSpot JVM. As a result, I’m also known to many as “Bull II”.

November 10, 2009 is my birthday, my previous works were hosted in Google’s codehub, and then more people used it, so the code was slowly migrated to GitHub. It’s also slowly catching on in China, so let’s give you a look at the scope of business I’m involved in:

In addition to large-scale projects such as Docker containers and kubernetes (known in the industry as k8s), which are implemented entirely in Go, more and more major Internet vendors are converting to Go. Today, the Go language has achieved a crucial position in the field of cloud computing and in the development of microservices for many large projects.

Why Go language is popular all over the world

• First, I’m a compiled language; I can execute the same program 30% to 70% faster than an interpreted language like Java;
• Secondly, as the C language of the 21st century, my development speed is not far behind, and my concise and efficient syntax leaves C and C++ developers in the dust;
• Concurrency support is friendly, and the advantage of being able to achieve concurrency in one line of code will make me more and more popular in high concurrency scenarios;
• Finally, the arrival of the cloud computing era and the massive demand for Internet servers will allow me to make a big splash in the future.

Don’t ask what the big picture is, ask is to surpass Java, python and other languages, and be among the top of the popularity.

Go, object-oriented?

Many of my partners who are new to Go are puzzled by the fact that I support encapsulation, but not inheritance or polymorphism, so strictly speaking I am not an object-oriented language. But I support interfaces to apply any kind of data type, and allow object-oriented programming style, from this level I am also a kind of object-oriented language.

So, am I an object-oriented language or not? Let’s take a look at how the official documentation defines me:

Answer: Yes and no. On the one hand Go has an object-oriented programming style, on the other hand it doesn’t have object inheritance, so Go is both an object-oriented language and not quite.
And it’s the lack of object inheritance that makes Go lighter and simpler than C++/Java. Think about base classes, derived classes, single inheritance, multiple inheritance, permission control, etc.

Quick start

1. hello, world First Go program

Whether you’re following me into the world of the Go language based on a job requirement, a desire to learn a highly promising development language, or your first foray into computers and programming.

After hearing so many stories about the king’s wife, I’m sure you’re familiar with this pushy routine. So, without further ado. Life is short, Let’s Go !

The first program in the life of the vast majority of computer professionals:

1
2
3
4
5
6

package main
import "fmt"
func main() {
    fmt.Println("hello world!")
}

As in many languages, the main function is the first function to be executed, but the main package is special in Go and is used to define a separate executable program. Simply put, the main package must be introduced wherever there is a main function and you need to run that main function:

1

package main

The fmt package contains a number of standard output and input functions, such as the print function Print, Println in the program means that immediately after printing a new line.

Don’t know how to run a Go program yet? Check out this article on installing and configuring Go to run

The Go language provides two common ways to run programs:* Build a .exe file with the go build command, and then run

go build hello.go This creates an .exe compiled file in the directory, which can be opened on a Windows machine, or executed from the command line or git bash.
/hello.exe /hello.exe

• Run the file directly with the go run command

go run hello.go

The result

hello world!

2. Find duplicate lines in a list of strings.

This program will teach you about for loops in Go, and the definition and use of string slices (slices: understood as variable length arrays):

1
2
3
4
5
6
7
8
9
10
11
12
13
14

package main
import "fmt"
func main() {
    arr := []string{"qq", "wechat", "dnf", "lol", "wechat"}
    counts := make(map[string]int)
    for i := 0; i < len(arr); i++ {
    counts[arr[i]]++
}
    for key, value := range counts {
    if n > 1 {
    fmt.Printf("值为 %s 的字符串重复了%d次 \n", key, value)
    }
    }
}

Define a string slice: arr := []string{}, with the variable name on the left, and := is a common assignment statement in Go, e.g.: i := 0, which is the equivalent of initializing the i variable before assigning it:

1

var i int i = 0

In my Go world, defining a map is usually done by initializing the map with make to prevent null pointer exceptions. Or you can define a map with initialized key-value pair elements when you create it:

1
2
3
4

counts := map[string]int {
    "qq" : 1,
    "wechat" : 2,
}

map [string] int, indicating a single string string as the key and the number of strings int type as the value.

Then, count the number of each string in the list of strings, here we have used the self-incrementing counts[arr[i]]++ to count the number of strings, equivalent:

1

counts[arr[i]] = counts[arr[i]] + 1

The advantage of this self-incrementation is that it eliminates the need for recurring variables and simplifies the code.

Finally, we iterate to print out the strings that have a number of occurrences greater than 1.

Run the result:

String with value wechat repeated 2 times

for is the only way to loop inside go, and there are two common forms:

1
2
3

for init; condition; post {

}

The for loop is formed without parentheses, but the curly braces after the for statement are required, and the left curly brace must be on the same line as the post statement.

1
2
3

for i, v := range array {

}

for range is used to quickly traverse data structures (arrays, slices, maps, etc.) with multiple elements, i is the index of the element (ranging from 0 to len(array)-1), and v is the value of the element.
That is, v == array[i]

Summary

This article introduces the history of the Go language and its general application scenarios. It’s worth mentioning that distributed development and microservices architecture are still at the cutting edge, and many programmers are still in traditional companies and small computer enterprises, and can’t get in touch with new technologies and projects.

So I wrote Go this series, but also want to let more want to learn but no way and direction of the computer direction of the school personnel, or has embarked on the work of the practitioners can really understand the charm of Go. In many large Internet companies, such as Microsoft, Byte, Tencent, Kingsoft, Xunlei, including Huawei such as communications companies, in many business teams have begun to Go transition, Go will also become the hottest development language in the era of cloud computing!

While the dawn is still uncertain, while the moon is still hanging in the sky, take your backpack and set out with me!

As developers, whether it is a job interview, or in the daily work, I believe that we are no stranger to high concurrency and multi-threading.

During job interviews, the backend job requirements that roll out of the sky often require us to familiarize ourselves with high concurrency and multi-process/multi-threading:

In our daily work, with the rise and development of mobile Internet applications, the system tasks and problems we face are becoming more and more complex.

Whether we are building large-scale web applications, processing huge data sets, or developing high-performance games, we all need to deal with a common challenge: high concurrency.

1.1 What is high concurrency?

** High concurrency refers to a large number of users or programs accessing and using a service or resource at the same time period. **

This means that we need to handle a large number of requests, data, and tasks at the same time. How to handle this situation efficiently becomes a critical technical task.

High concurrency is an area of challenge, but also an area of opportunity.

1.2 What does multithreading have to do with high concurrency?

Solving the problem of high concurrency not only improves the performance of the system, but also improves the user experience and brings more business opportunities for the enterprise.

Multi-threading technology is one of the most important tools to address the challenge of high concurrency. **

Therefore, in this post, Xiao ❤ will take you together to explore high concurrency and multithreading in depth, and familiarize you with the working principle of multithreading, application scenarios, as well as practical solutions to solve the problem of high concurrency.

I believe that no matter you are a junior programmer or a developer with some experience, you will be able to find useful information in this article.

2. High Concurrency

2.1 Concurrency and Parallelism

Concurrency

Concurrency is the execution of multiple tasks in the same time period. On a single-core processor, multiple threads switch execution between themselves by means of time-slice rotation, resulting in a concurrent scenario.

Equivalent to our video recordings, when the video frame rate is high enough (i.e., switching between multiple images in a second), our naked eye treats it as a continuous and smooth video.

Parallel

On a multicore processor, true concurrency can be achieved by multiple threads performing different tasks at the same time - that is, in parallel.

Parallelism is the execution of multiple tasks at the same moment, usually requiring a multi-core processor. Parallelism is a subset of concurrency, and true parallelism can only be achieved if the hardware supports multiple parallel execution units.

Thinking Questions

Scenario 1: Why is it difficult for us to focus on answering the phone while playing a game with an intense group battle?

Scenario 2: We can listen to music while fiddling with the steering wheel without interfering with each other while driving. Why don’t you guess whether our brains run concurrently or in parallel?

2.2 How much concurrency is too much concurrency?

Having understood the concept of concurrency, let’s now talk about high concurrency.

High concurrency is a relative concept that depends on the performance and processing power of the system. ** Typically, a system is said to be highly concurrent when the number of requests or transactions it needs to handle exceeds its normal load. **

2.3 Challenges of High Concurrency

While high concurrency brings many opportunities, it also comes with many challenges.

For example, a highly concurrent system needs to handle a large number of requests in a short period of time without degrading the performance or responsiveness of the system.

This may involve multiple users accessing a website at the same time, multiple clients requesting server data at the same time, or multiple threads accessing shared resources at the same time.

In a distributed system, whether it’s multiple users accessing, or multiple clients accessing a server, it all boils down to the business threads of each server accessing shared resources, so **high concurrency challenges are almost always related to multithreading. **

When faced with high concurrency, the following problems specifically arise.

1. Competing conditions

Multiple threads accessing shared resources at the same time may lead to data inconsistency problems. For example, multiple threads depositing money into the same bank account at the same time may result in an incorrect balance.

2. Deadlock

Multiple threads wait for each other to release resources, causing the system to stall. For example, thread A waits for thread B to release a lock, and thread B waits for thread A to release a lock, creating a deadlock.

3. Resource Contention

Multi-threaded access to shared resources can lead to resource contention problems that can degrade performance. For example, multiple threads competing for a database connection at the same time results in a slower database response.

4. Thread Safety

There is a need to ensure that no errors are raised when multiple threads access shared data. For example, in a multi-threaded environment, there is a need to ensure that reading and writing to data is safe.

5. Debugging Difficulty

Since the order of execution of multithreading is uncertain, problems may appear at different times. So debugging of multithreaded programs is relatively complex and problems are difficult to reproduce.

2.4 Solving High Concurrency Problems

In order to solve the high concurrency problem, appropriate techniques and methods need to be used, which are as follows.

1. Lock Mechanism

Locks are used to protect shared resources by ensuring that only one thread can access them at a time.

Locks can be categorized as 互斥锁 and 读写锁. Mutual-exclusion locks are used to exclusively occupy a resource, while read/write locks allow more than one thread to read a resource at the same time, but only one thread is allowed to write to it.

Specific implementation details can be found using the locking mechanisms provided by the programming language, such as the synchronized keyword in Java or threading.Lock in Python.

2. Concurrent Data Structures

Use concurrent data structures such as concurrent queues and hash tables to reduce resource contention. These data structures are optimized to work efficiently in a multithreaded environment.

For example, Java provides ConcurrentHashMap, which is a thread-safe hash table that can be used in highly concurrent environments without explicit locking.

3. Thread Pooling

Manage and reuse threads to improve performance. A thread pool controls the number of threads, to avoid wasting resources by having too many threads.

In Java, you can use ExecutorService to create and manage thread pools. This avoids frequent creation and destruction of threads and improves efficiency.

4. Message Passing

Inter-thread communication through message passing model avoids shared memory. Message passing ensures that data is passed safely between threads, reducing competing conditions.

For example, in Go, you can use channels (channel) for messaging to ensure the safe passing of data.

5. Atomic operations

Atomic operations are indivisible operations that ensure that multiple threads operating on shared variables are safe. Atomic operations are often supported by third-party libraries or features that can be used to implement various synchronization mechanisms.

In C/C++, you can use atomic operations to manipulate shared variables, for example with the atomic library. In MySQL, the transaction threads of the InnoDB engine can carry their own atomicity features.

3. Multithreading

3.1 Processes and Threads

When a task in concurrent work is completed, it will be switched from one segment of the program to another to be executed, and a series of states of the previous segment of the program operation will be lost if not saved, so the operating system introduces processes to carry out resource isolation.

1. Processes

Processes are used to delineate the basic unit of resources required when a program runs, it has an independent address space, an independent stack, when the process is switched, it can ensure that the respective data storage is not affected.

Because the process involves the consumption of a large number of resources, it is strictly controlled by the computer operating system (can be understood as: the approval of land resources in each province and city, are very cautious, especially the first-tier cities, so by the core department of unified control).

Therefore, ** process switching occurs in the kernel state, by the computer core program to unify the scheduling. **

Trivia:
The operating system is divided into kernel state and user state, and the CPU (Central Processing Unit) in the kernel state can access arbitrary data.
The CPU in the kernel state can access any data, > including peripheral devices such as network cards and hard disks, and there is no preemption of the occupied CPU.
Whereas a CPU in the user state can only access memory in a restricted manner and is not allowed to access peripheral devices, and the CPU in the user state may be seized by other programs.

2. Threads

When a process is switched, the kernel state has to be switched, so it consumes a lot of resources, for which the concept of threads is introduced.

A thread is the smallest unit of operating system scheduling and is an execution process within a program. A process can contain multiple threads, which share process resources such as memory space and file handles, but each has a separate stack memory.

The thread itself takes up almost no resources, it ** shares address space and heap ** with other threads in the process, so the scheduling time consumption is relatively small, but it has an independent CPU context (including CPU registers, program counters, etc.).

Threads are like sharing the same land resources with the threads inside the same process, but threads have their own office buildings, and switching between threads is unified scheduling by the operating system.

Trivia: Threads are divided into kernel threads and user threads, and user threads must be bound to kernel threads before they can run.

3.2 Multithreading Concepts

Multithreading is a type of concurrent execution that allows a program to be divided into multiple independent threads, each of which can perform tasks independently. It’s like being able to work on a piece of land resource at the same time without affecting each other.

1. Creating and managing threads

Creating and managing threads involves the scheduling mechanism of the operating system, which is implemented in different ways in different programming languages. Let’s take Python as an example:

1
2
3
4
5
6
7

import threading

def my_function():
  

thread = threading.Thread(target=my_function)
thread.start()  

2. Thread synchronization and mutual exclusion

When multiple threads access a shared resource at the same time **, it may lead to a race condition **, where multiple threads compete with each other for the resource, which may result in inconsistent data.

To solve this problem, we use a locking mechanism to ensure that only one thread can access the shared resource at the same time.

1
2
3
4
5
6
7
8

import threading

lock = threading.Lock()

def my_function():
   lock.acquire()  

   lock.release() 

3.3 Multi-threaded applications

Multi-threading can not only improve the performance of the program, but also improve the user experience. In real life, we often encounter multithreaded application scenarios.

1. Web server

Imagine a popular social media site with millions of users visiting the site at different times at the same time.

These users request different pages, upload photos, make posts, and also have some background tasks such as data backup, new post push, and so on.

At this point, the **Web server needs to handle requests from multiple users at the same time. Each user request can be seen as a thread, and multithreading allows the server to respond to multiple requests at the same time. **

For example, one user may request to view their profile, while another user may request to post a new status update. These two requests can be handled by different threads at the same time, improving the server’s response time.

2. Database system

Suppose an online banking system where thousands of customers access their account information simultaneously, checking balances, transferring funds, and so on. In addition, the banking system needs to handle customers’ deposit and withdrawal operations.

At this point, the database system needs to handle multiple customer requests at the same time. Each customer request can be regarded as a thread, and multiple threads can query the database at the same time to ensure that each customer’s account information is up-to-date.

3. Game Interaction

A multiplayer online game where dozens of players participate in the game at the same time. This game needs to handle player actions, physics simulation, AI computation, and multiplayer game interaction simultaneously.

At this point, the game engine can use multiple threads to handle different aspects of the task. One thread can be responsible for rendering the game graphics, another thread can handle the player’s actions, and another thread can be responsible for simulating the physics effects in the game.

With multi-threading, the feedback of the game system will be smoother and the player can enjoy a highly interactive game experience.

4. Summary

When talking about multithreading and concurrency, it’s like the busy streets of our daily lives, where everyone is dealing with their own things, but at the same time need to coordinate their interactions with others.

1. Introduction
2. why microservices are needed
3. Service Discovery
4. Inter-Service Communication
5. Microservice Splitting

1. Introduction

1.1 The Status of Microservices

In the current Internet world, no matter whether you are doing 2B-end [for enterprises] or 2C-end [for individual users] products, microservices development has blossomed everywhere, and most of the monolithic structure of the system only exists in the traditional software and medium-to-large-scale state-owned enterprise architectures.

But do you really understand microservices?

Let’s take a look at the requirements for microservices in a backend development engineer job JD:

**Candidates need to be proficient in the use of microservices frameworks, deeply understand the principles and operation mechanisms of microservices, and have unique insights into service splitting, inter-service calls, and service governance. **

Is it feeling cloudy or half-understood? It’s okay, we’ll get it after reading this article 😉

1.2 What are Microservices

There are many definitions of microservices, the official one is as follows:

Microservices are an architectural and organizational approach to developing software that consists of small, independent services communicating through well-defined APIs that are handled by small, independent teams. Microservices architecture makes applications easier to scale and faster to develop, thereby accelerating innovation and reducing time to market for new features.

In layman’s terms, microservices are an architectural pattern, or an architectural style. It advocates the division of a single application into a number of smaller services, each of which is managed independently, with services and services cooperating and coordinating with each other, ultimately providing users with a rapidly iterative go-live product.

For example, an e-commerce system may contain micro-services such as users, commodities, warehousing, orders, payments, and points systems:

Today we will talk about microservices in detail and the knowledge related to microservices interview. I believe that after reading this article, no matter whether you are dealing with an interview with a big internet company or a state-owned foreign enterprise, microservices are earned when asked 💰

2. Why we need microservices

2.1 The problem with monolithic architectures

In the beginning, when there were no microservices, applications in the Internet world used a monolithic architecture, i.e., all the modules under the system were put into a single service.

Often, these services are as small as a few, or as large as dozens of modules, and teams of dozens of people work together to develop them. At this time, the problems of monolithic architecture are gradually exposed:

1. systems are usually accessed by APIs, and the tight coupling makes it difficult to maintain and expand;
2. each business area needs to use the same technology stack, it is difficult to quickly apply new technologies, such as using Java or PHP;
3. System modifications must be deployed/upgraded together with the whole system, which makes operation and maintenance complex and error-prone;
4. When the system load increases, it is difficult to scale horizontally;
5. If there is a problem in one part of the system, it may affect the whole system and cause collateral problems.

Thus, in 2012, the concept of microservices was proposed as a way to accelerate the development of Web and mobile applications, and began to attract much attention.

Time came to 2015, more and more Internet communities, forums, and Internet giants began to use microservices, and from 2018, more and more small and medium-sized enterprises also began to upgrade their microservices architecture.

Today, in 2023, according to a recent market research report, 95% of the applications in the market have been developed using microservices architecture.

2.2 Disadvantages of Microservices

1) High code complexity

Microservices interact with each other via HTTP, RPC, etc. Compared with the API form of monolithic architecture, it is necessary to take into account the failure of the called party, overloading, loss of messages, etc., and the code logic is more complex.

Transactional operations between microservices need to solve the problem of distributed transactions, and may need to introduce two-stage, best-effort notification and other ideas to solve the problem.

When there are a few functions (or database fields) overlapping between microservices, but they cannot be extracted into microservices, it is usually necessary to repeat the development, or do data redundancy, which increases the development and maintenance costs.

2) Heavy operation and maintenance tasks

At the time of launch, there may be service coupling relationships that require sequential and orderly service deployment. The most common situation is that A service calls the interface of B service, i.e., A depends on B. When you go online, you have to upgrade the B system first and then the A system.

Moreover, since the system is independent, a well-designed monitoring system is needed to monitor the operation status of each microservice. The purpose of real-time monitoring is to prevent modules in the business system from dropping out temporarily, which affects the overall business utilization.

3) Impact on performance

Compared to the monolithic architecture, the delay of REST and RPC communication between microservices is higher, especially when the call chain is long.

Moreover, when the business chain is long, it is more difficult to troubleshoot problems, and if multiple services are involved in the BUG, troubleshooting will be more complicated.

• 2.2 Characteristics & Principles of Microservices

  Although microservices have disadvantages that do not exist in monolithic architectures, its advantageous characteristics of allowing high cohesion and low coupling of functional modules are enough for us to endure these to embrace it. In addition to this, microservices have the following characteristics:

  1) Single responsibility

  Generally, microservices are divided according to business logic, and each microservice is only responsible for the functions belonging to its own business domain, with clear logic and high module cohesion. For example, the user system, product system, order system, payment system, etc. mentioned above, each module has its own business characteristics.

  2) Autonomy

  Microservices are independent entities that can be deployed and upgraded alone, and each microservice communicates with each other through standard interfaces in the form of REST/RPC. And microservices can be realized with different technology stacks, and other modules are not affected.

  Most commonly, many algorithm modules are implemented in Python, and the background business is implemented in Go or Java When the background module needs to call the interface of the algorithm module, as long as the protocol of the API interface is defined, the API call can be made in a friendly way.

  Through isolation, meltdown and other techniques to ensure that a microservice exception does not affect other modules, which is also a microservice autonomy of the bottom of the pocket strategy.

  3) Scalable

  With business growth, a module can be expanded horizontally or vertically to facilitate elastic scaling and gray-scale release.

  In different business scenarios, the pressure of different modules is inconsistent, for example, the e-commerce system under the limited-time spike business: the commodity system and the storage system will bear most of the traffic pressure, while the payment system may be more relaxed.

  In order to cope with similar scenarios, we can distribute the modules of the merchandise system and the storage system to multiple machine deployments, and equally distribute the traffic pressure to different business machines to alleviate the high concurrency pressure of the modules.

  4) Flexible Combination

  Under the microservice architecture, functional reuse can be achieved by combining existing microservices. Although this can also be realized under the monolithic architecture, different combinations may make the interactions between monolithic architecture modules more and more complex, and ultimately make the system chaotic.

  3. Service Discovery

  3.1 Service Governance

  As applications evolve from monolithic architectures to microservices, the demand for service governance such as service discovery, load balancing, and fusion-limiting between microservices increases significantly due to the substantial growth in the number of fine-grained microservice applications.

  In a microservice scenario, each service has multiple instances, and a mechanism is needed to resolve the requested service name to the corresponding service instance address, which requires service discovery and load balancing mechanisms.

  3.2 Service Discovery

  Service discovery consists of two parts:

  • Service registration: each service name sends information about the service instance to the registry and provides a heartbeat mechanism to ensure that the service is online;
  • Service discovery: get the list of instances corresponding to the service from the registry.

3 common strategies for service discovery:

1. ETCD as a registry, each service module registers the service name and instance information to ETCD, and keeps its own lease unexpired to maintain the service online;
2. Redis as the registry, each module maintains a timer and sends instance information and time to the registry module periodically, and the registry module derives the maximum timeout time based on the heartbeat time and network fluctuation time. If a module does not send the heartbeat message packet to the registry for a long time, it will be kicked out of the service cluster;
3. Istio service registration and discovery is accomplished by the control plane Pilot and the data plane Envoy, the Pilot obtains service resource information such as service and endpoint through the K8s APIServer interface, and converts it into xDS messages to be sent to the Envoy component in the data plane, and the Envoy selects a service instance to be requested according to the load balancing policy configured when it receives the request. When the Envoy receives the request, it selects a service instance to forward the request according to the configured load balancing policy.

3.3. Load Balancing

1) What is load balancing?

Load balancing is generally used in conjunction with service discovery to select one of the service discovery parsed instances to initiate a request, and this process uses a load balancing policy.

As shown in the figure above, when the business deploys 4 commodity system services, when the request arrives at the registration center, **the registration center will hit the request to different servers **according to different load balancing algorithms to ensure that the load on each machine is not too high to affect performance.

2) Load Balancing Algorithm

Compare balanced forwarding to a teacher handing out candy, where the teacher is the registry and the students are the servers. Load balancing algorithm is a way for the teacher to give out candy to the students, neither letting the students starve nor giving candy to only one student [it’s not good to have too much blood sugar].

1, polling method

Assigns requests to back-end servers in sequential rotation, equalizing each back-end server, doesn’t care about the number of connections to the server or the current system load.

Using the analogy of a teacher handing out candy, the polling method is to hand out candy to students one by one, student 1, student 2, student 3 …… The rain falls evenly, and each student gets an almost equal number of candies.

2, randomized method

Randomly select a machine to visit, by the probability statistics know, When the number of calls, the more its allocation calls closer to the average, that is, the result of polling.

Using the analogy of a teacher handing out candy, the random method is that the teacher draws lots to decide who to give the candy to each time before handing it out, and each student gets an indeterminate number of candies. But probabilistically, when there are a lot of requests, the result is close to the average, similar to rolling dice, or tossing a coin.

3, weighted polling method

According to the hardware configuration and load of the server, the machine with high performance and low load is given a higher weight, making the machine with high weight more accessible.

When the number of requests is large, the ratio of the number of requests handled by each service converges to the ratio of weights.

Using the analogy of a teacher handing out candy, the weighted polling method looks at the teacher’s preferences for handing out candy, where the weights can be thought of as the student’s score. The teacher might give 3 candies to students who score 90 or more, 2 candies to students who score 75 or more, 1 candy to students who score 60 or more, and no candy to students who fail.

The ratio of the number of candies each student receives is based on the weights, so when the number of requests is higher, the ratio of the number of candies each student receives to the weights is closer.

4, source address hashing method

Get the IP address of the client and access a particular server by hash taking the mode. When the list of back-end servers and the hash algorithm are unchanged, the same client is mapped to the same server for each request.

To use the analogy of a teacher handing out candy, the source address hash method means that the teacher hands out candies of the same origin or brand to one student at a time, e.g., Heijirou candies are given to student 1, Shanghaojia candies are given to student 2, and so on. …… The number of candies each student gets depends on the number of candies of the corresponding brand.

Translated with www.DeepL.com/Translator (free version)

• 4. Inter-service calls

  We have already introduced the characteristics of microservices above: ** Microservices interact with each other through HTTP, RPC, etc. **.

  HTTP (HyperText Transfer Protocol) and RPC (Remote Procedure Call), the former is a protocol and the latter is a method, both of which are commonly used for service invocation.

  RPC works on top of the TCP protocol (or HTTP), while HTTP works on top of the HTTP protocol, which is based on the TCP transport layer protocol, so RPC is naturally lighter and more efficient than HTTP.

  4.1 HTTP

  1) Introduction

  HTTP service development, i.e. developing RESTful style service interfaces, is a common means of communication for information transfer when there are few interfaces and little interaction between systems.

  2) Advantages

  The advantages of HTTP interfaces are simple, direct, easy to develop, and can be transported using the ready-made HTTP protocol. When developing a service, it is necessary to agree on an interface document that strictly defines the inputs and outputs and specifies the request methods and parameters of the interface.

  4.2 RPC

  1) Introduction

  **First of all, we need to know what is RPC? **

  RPC (Remote Procedure Call) is a computer communication protocol. The protocol allows a program running on one computer to call a subroutine in another address space (usually a computer on an open network), and the programmer calls it as if it were a local program, without having to additionally program for this interaction (no need to pay attention to the details). –Wikipedia

  A lot of non-specialists must have been confused when they saw this big explanation, as I was at first. It’s not hard to understand RPC, we just need to know that it’s a communication protocol, i.e. a format or convention that both parties need to follow.

  2) Characteristics

  Some people may still wonder: since they are both communication protocols, should we choose HTTP (HyperText Transfer Protocol) or RPC for program interaction and application development?

  This starts from the difference of their properties, which are mainly considered from the following 4 points:

  transfer protocol

  • RPC is a communication protocol based on the TCP transport layer or the HTTP2 application layer;
  • HTTP is based on HTTP protocol only, including HTTP1.x (i.e. HTTP1.0, 1.1) and HTTP2, and many browsers currently use 1.x by default to access server data.

  Performance Consumption (from data type comparison)

  • RPC, can be based on gRPC (a RPC framework) to achieve efficient binary transmission;
  • HTTP, mostly implemented via json, where byte size and serialization consume more performance than gRPC.

  load balancing

  • RPC, basically comes with a load balancing strategy;
  • HTTP, need to configure Nginx, HAProxy to realize.

  Transport Efficiency

  • RPC, use customized TCP protocol, can make the request message size smaller, or use HTTP2 protocol, can also be very good to reduce the size of the message, improve the transmission efficiency;
  • HTTP, if it is based on HTTP1.x protocol, the request will contain a lot of useless content; if it is based on HTTP2.0, then a simple encapsulation can be used as RPC, which is when the standard RPC framework has more advantages of service governance.

  3) Popular RPC Frameworks

  • gRPC: based on the HTTP2.0 protocol, the underlying use of the Netty framework;
  • Thrift: cross-language service development framework, through the code generator to save a series of basic development work;
  • Dubbo: protocol and serialization framework can be plugged and unplugged.

  Translated with www.DeepL.com/Translator (free version)

4.3 Differences between HTTP and RPC

In summary, we can easily find that RPC is stronger than HTTP in terms of performance consumption and transmission efficiency, as well as load balancing. At this point, careful friends may have found, then why our common systems and websites are using the HTTP protocol, not changed to RPC communication?

To give a common example, HTTP is like Mandarin, RPC is like a local dialect, such as Cantonese, southwest of Yunnan, Guizhou, Sichuan.

The advantage of speaking Mandarin is that everyone understands it, and most people speak it, so HTTP has a certain universality. Speaking dialect, the advantage is that it can be more concise, more confidential, more customizable, the disadvantage is that the other party who “speaks” the dialect (especially the client side) must also understand, and once everyone speaks a dialect, it is difficult to change the dialect.

So RPC is generally used for internal service calls, such as between service A and service B in Alibaba’s Taobao system.

The concept of microservices emphasizes more on independence, autonomy and flexibility, but RPC has more restrictions. Therefore ** microservice frameworks generally use HTTP RESTful calls, except for systems with high efficiency requirements **.

5. Service Splitting

Through the above description, we must have familiarized ourselves with the basic features of microservices from several levels of microservice introduction, service governance, service discovery and inter-service communication.

** Then as an architect/intermediate and advanced programmer, how do we go about distinguishing the boundary issues of each module and how to split microservices? **

Next, we introduce three common ways to split microservices.

5.1 Business Domain

Splitting microservices by business domain (also called pendant splitting), such as user, mall, order business modules, if there is the same functionality needs to be aggregated, it is carried out to sink to a separate microservice, unified call.

The benefits of splitting microservices by business are, high cohesion and no easy coupling between businesses.

5.2 Functional Positioning

Split microservices by functional positioning (also called horizontal splitting), such as login and registration, user shopping, and points redemption. If the same module is used by more than one function, you can further split the microservices and call them uniformly.

The benefits of splitting microservices by function localization are: ** high development efficiency, more independent testing and use of different functions**.

For users, when using a certain type of system function, they often call the API of the same microservice, which can better avoid collateral problems and distributed transaction problems.

5.3 Level of Importance

Split microservices according to the degree of importance, distinguishing between core and non-core modules, e.g.: e-commerce system inside the order module core, logistics module non-core. The key points to distinguish whether a module is core or not are:

1. Is it indispensable. Such as the e-commerce system, the user’s most concerned about the function of online shopping, orders is to save the user’s shopping records of the core business, naturally, is very important.
2. High user attention, i.e., high traffic, can bring exposure for the system. For such a module (e.g., e-commerce display module), it will greatly affect the user’s judgment of the overall product quality, which is also extremely important.

In addition, the industry on the topic of microservice splitting has a relatively well-established system design methodology, such as the famous DDD (Domain-driven design) .

Its principle is to establish domain model through event storm, reasonably divide domain logical and physical boundaries, establish domain object and service matrix and service architecture diagram, define code structure model that conforms to the idea of DDD layered architecture, and ensure the consistency of business model and code model.

Due to personal understanding and the limited length of the article, I will not expand on the discussion here. If you want to understand the friends move a small hand point a concern, the follow-up time can be arranged.

Translated with www.DeepL.com/Translator (free version)

Today, I’m going to talk to you about refactoring, the mysterious skill of programmers! Don’t worry, I’m going to help you understand and master this skill with easy-to-understand language and some fun conversations that my 8-year-old niece says she understands.

1.1 Background

Code Development:

One month later:

Later there’s time change it (don’t worry, there won’t be time for that, and I won’t change it when there is time).

Six months later:

As above, it is a scenario that any developer will experience: the early code simply cannot be reviewed, or else it will surely fall into deep suspicion, is such a bad code really from their own hands?

What’s more, most of the current systems are collaborative development, and each programmer’s naming conventions and coding habits are different, leading to a situation of one system code, multiple flavors.

What is refactoring

Yeon: Hey uncle, I heard you’re going to share refactoring, what’s new about that?

❤：Hi Yeon! Refactoring is the process of improving the design of existing code to make it more understandable and easier to maintain, without changing its functionality. Think of it as giving your code a beauty makeover, but on the inside it’s still the same code, not an “unrecognizable person”.

Why Refactor

Lulu: Wow, that sounds awesome, why do we need to refactor?

❤：Haha, good question, Lulu! Because code is alive and getting bigger day by day, when it becomes difficult to understand and modify, it is like a heavy elephant that slows down our pace. Refactoring is like putting weight on the elephant, making it lighter and more flexible, and development speed can be increased quite a bit!

In the same way that you guys have a little cleanliness fetish and love to clean up your room, programmers with code cleanliness fetishes refactor their code all the time!

When to refactor

Yeon: Sounds reasonable, but when should you use refactoring?

❤：Good question, Yeon! There are several scenarios:

• When you see several places in your code that look exactly the same, this is a good time to consider merging them into one to reduce redundancy.
• When you have a function or method that looks thicker than a dictionary, break it up into smaller parts that are better understood.
• When you want to fix a bug, but realize that the original code structure is so complex that fixing it becomes as hard as solving a puzzle, refactoring before fixing it is a good idea.
• When you want to add new functionality, but the code doesn’t let you extend it easily, it’s also a good idea to refactor first and then extend.

Steps to refactoring

Lulu: I see uncle, so what are the exact steps of refactoring?

❤：Good question, Lulu, it looks like you’re seriously thinking about it! Let me walk you through the basic steps of refactoring next!

2. How to refactor

Before refactoring, we need to identify the bad-flavored code inside the code.

By bad flavor, I mean the superficial messiness of the code, and the deep-seated corruption. Simply put, it’s code that just doesn’t feel right.

2.1 Bad-flavored code

In the book Refactoring-Improving the Design of Existing Code, these two dozen bad taste situations are described, and we’ll pick the most common ones below.

1) Methods that are too long

Methods that are too long are those that do too much work in a single method, often accompanied by statements in the method that are not at the same abstraction level, e.g., a mix of dto and service level code, i.e., fragmented logic.

In addition to this, methods that are too long tend to cause a number of additional problems.

Problem 1: Excessive annotations

Methods that are too long are difficult to understand and require a lot of comments. If 10 lines of code require 20 lines of comments, the code is hard to read. If 10 lines of code require 20 lines of comments, the code is hard to read. Especially when reading code, you often need to remember a lot of context.

Problem 2: Procedure-Oriented

The problem with process-oriented code is that when the logic is complex, the code can be difficult to maintain.

On the contrary, we often use object-oriented design thinking in code development, i.e., abstracting things into objects with common characteristics.

Solution Ideas

To solve the problem of long methods, we follow the principle that whenever we feel the need to write a comment to explain the code, we write this part of the code into a separate method and name it according to the intent of this code.

Method naming principle: it can summarize what to do, not how to do it.

2) Overly large classes

A class that does too many things, for example a class whose implementation contains both product logic and order logic. There will be too many instance variables and methods that are difficult to manage at creation time.

In addition to this, too large a class tends to cause two problems.

Problem 1: Redundancy and duplication

When a class contains the logic of two modules inside, the two modules are prone to create dependencies. This can easily lead to the problem of “you take me, I look at you” during the code writing process.

That is, in both modules, the program structure or methods with the same intent are seen related to the other module.

Problem 2: Poor coupling structure

When the naming of a class is insufficient to describe what it is doing, the probability is that the coupling is poorly structured, which is contrary to the goal of writing code with “high cohesion and low coupling”.

Solution

Split large classes into small classes based on business logic, and if there is a dependency between two classes, associate them through foreign keys and so on. When duplicate code occurs, try to merge and present it as much as possible, and the program will become more concise and maintainable.

3) Logical decentralization

Logical fragmentation is due to unreasonable dependencies at the code architecture level or object level, which usually leads to two problems:

Fragmentation

A class is often modified in different directions for different reasons.

Scattered changes

When a certain change occurs, it needs to be modified in multiple classes.

4) Other bad flavors

Data Mud Clusters

A data mudball is a confusing blend of many data items that are not easily reused and extended.

It is easier to categorize many data items when they always appear together and when they appear together. We can then consider encapsulating data into data objects by business. The reverse example is below:

1

func AddUser(age int, gender, firstName, lastName string) {}

After the refactoring:

1
2
3
4
5
6
7

type AddUserRequest struct {
   Age int
   Gender string
   FirstName string
   LastName string
}
func AddUser(req AddUserRequest) {}

Basic Type Paranoia

In most high-level programming languages, there are basic types and structural types. In Go, basic types are int, string, bool, and so on.

Basic type paranoia refers to the fact that when defining variables for an object, we often use the basic type without considering the actual business meaning of the variable.

An example is the following:

1
2
3
4

type QueryMessage struct {
Role        int         `json:"role"`
Content  string    `json:"content"`
}

After the refactoring:

1
2
3
4
5
6
7
8
9
10
11
12

type MessageRole int

const (
HUMAN     MessageRole = 0
ASSISTANT MessageRole = 1
)

type QueryMessage struct {
Role        MessageRole   `json:"role"`
Content  string               `json:"content"`
}

• This is the request field when ChatGPT asks a question, we can see that the conversation role is of type int, and 0 means human, 1 means chat assistant.

  When we use int to represent the conversation role, there is no way to know more information directly from the definition.

  But by defining it with type MessageRole int, we can clearly see that there are two types of conversation roles based on constant values: HUMAN & ASSISTANT.

  Confusing Hierarchical Calls

  Our general system is hierarchical based on business service, transit controller and database access dao. Generally, the controller calls the service and the service calls the dao.

  If we call the dao directly from the controller, or if the dao calls the controller, there will be a hierarchical confusion, which can be optimized.

  5) Problems caused by bad flavors

  YanYan: Uncle, do all these bad flavors need to be addressed, and what kind of impact do you think these bad flavors of code will bring?

  ❤: Yes, code with too many bad flavor codes will bring four “difficult “ .

  • Difficult to understand: new developer students can’t understand the code of the people watching, a module after two weeks of reading still don’t know what it means. Maybe it’s not the level of the developer is not enough, may be the code is written too hard to say.

  • • • difficult to reuse * * *: either read can not read, or barely read but do not dare to use, for fear of what dark pit. Or the system is so coupled that it’s hard to separate the reusable parts.

  • Difficult to change: affects the whole body, i.e., scattershot modification. Move one piece of code and the whole module is almost gone.

  • Difficult to test: change is not good to test, it is difficult to carry out functional verification. The naming is messy and the structure is confusing, and new problems may be measured during testing.

  3. Refactoring Tips

  Lulu: Oh, so that’s it, can we remove them then?

  ❤: Of course you can! Just like you guys love to clean up your room, every responsible (code cleanliness) programmer will consider code refactoring.

  And the industry already has a better way of thinking about the refactoring problem: removing the “bad taste” from the code through continuous refactoring.

  1) Naming conventions

  A good naming convention should conform to:

  • Accurately describe what is done
  • Formatting that conforms to common conventions

  common conventions

  Let’s take Huawei’s internal Go language development specification as an example:

  Scenario Constraints Example Project names are all lowercase, separated by an underscore ‘-‘ for multiple words user-order Package names are all lowercase, separated by an underscore ‘-‘ for multiple words config-sit Structure names are initialized in uppercase Student interfaces use Restful API naming, with the last portion of the path being a resource noun such as [get] api/v1/ student constant name initial capitalization, camel naming CacheExpiredTime variable name initial lowercase, camel naming userName, password

  2) Refactoring techniques

  YanYan: Wow, so many mature specifications can be used ah! So besides the specification, do we need to pay attention to anything else?

  ❤: Good question Yeon! Next I will also introduce some common refactoring techniques:

  • Extract function: break a long function into smaller chunks, easier to understand and reuse.
  • Rename: rename variables, functions, classes, etc. to make more sense.
  • eliminate redundancy: find similar chunks of code and merge them to reduce duplication.
  • Move: move functions or fields to more appropriate places to make the code more organized.
  • Abstract generic classes: pull out generic functions into a class to increase code reusability.
  • Introduce parameter objects: when there are too many variables, pass in objects to eliminate data clumps.
  • Use the guard statement: reduce the use of else and make the code structure clearer.

4. Summary

Lulu: Uncle, you’re so funny, I feel like I can refactor too!

❤: Lulu is awesome, I believe in you! The idea of refactoring is everywhere, just like all life should be white space, your life will be very wonderful too. In programming, refactoring can make the code more beautiful, easier to read and understand, and improve the development efficiency, which is a skill that all programmers should master.

YanYan: I can do it, too! In the future, I will also write code and do code refactoring, and I will also like my uncle’s article.

❤：Hahaha, well da, you guys are great! Just like you guys like to clean and paint and read poetry, if you want to write code in the future, they’ll be very clean and poetic too!

As I was holding my cell phone the other day, chatting freely with my friends’ WeChat group about gossip news and upcoming weekend plans, suddenly a message with a joyful message came to me with eight big words written right in the middle of the message: Congratulations on your fortune, great luck.

Grab the red envelopes! I’m sure most of you are no strangers to this, so how is this group chat system of WeChat designed to make it easy for us to chat, share pictures and emoticons, and that magical red packet feature?

This question has been bothering me for a long time, so I decided to dig a little deeper and see how the design behind WeChat’s group chat system works.

WeChat Group Chat System Design

WeChat, as a universal App with 1 billion users, must have been used by all of you. The WeChat group building function is a core ability inside WeChat, which can put hundreds of friends or strangers into a group space.

Maybe you’ve experienced group chat on WeChat many times, but have you ever wondered how the system behind this is designed?

Let’s explore it today.

2. System Requirements

2.1 System Features and Functional Requirements

WeChat Group Chat is one of the core features of the social application, which allows users to create their own social circles to communicate with family members, friends, or enthusiasts of common interests in a friendly way.

The following are the core features of WeChat group chat system:

• Create Group Chat: Users can create new chat groups, invite other friend users to join or build groups with strangers face to face.
• Group Management: Group owners and administrators are able to manage group members, set rules and permissions.
• Message Sending and Receiving: Allows group members to send multiple types of messages such as text, image, audio, video, etc. and push them to all group members.
• Real-time communication: messages should be able to be delivered quickly to ensure real-time interaction.
• Red Packet Grabbing: Users can send any number and amount of red packets in the group chat, and group members can grab the red packets with random amount.

2.2 Non-functional Requirements: Coping with High Concurrency, High Performance, and Mass Storage

When we face the scenario that 1 billion WeChat users may use the group building function every day, we need to deal with large-scale user concurrency. This leads to the non-functional requirements of the system, including:

• High Concurrency: the system needs to support a large number of users creating and using groups simultaneously to ensure a latency-free user experience.
• High performance: fast messaging and instant response are key to digital socialization.
• Massive Storage: The system must be scalable to accommodate massive amounts of user-generated message text, images, and audio/video data.

3. Outline Design

In the outline design, we consider the outline design of the core components and basic business of the system.

3.1 Core Components

The following core components and protocols are involved in the WeChat group chat system.

• Client **: Receive messages from WeChat group chat on cell phone or PC and transmit them to the backend server in real time.

• Websocket transfer protocol: supports real-time interaction between the client and the backend server, with low overhead and high real-time performance, commonly used in WeChat, QQ and other IM systems communication systems

• Long-connection cluster: a cluster of systems that make long Websocket connections with clients and forward messages to application servers through middleware.

• Message Processing Server Cluster: provides real-time message processing capability, including data storage, query, and interaction with the database.

• Message Push Server Cluster: this is the relay station for messages and is responsible for delivering messages to the correct cluster members

• Database server cluster: it is used to store user text data, thumbnails of images, audio and video metadata, etc.

• Distributed file storage cluster: for storing user’s pictures, audio and video files.

3.2 Business Outline Design

Group Chat Creation

• Unique ID Assignment: When a user requests to create a new group, the system generates a unique group ID, which can be generated by a distributed ID generator such as Snowflake, or by using a database incremental ID. here, we adopt MySQL’s incremental ID for the sake of simplicity.
• Group Information Storage: Stores group IDs and related information (e.g. group name, creator ID, etc.) in the group database.
• Member association: add the group owner as the founding member of the group, and the creator will also become the administrator.
• Message History: To ensure that new members can access previous messages, the group ID of this new group is stored in association with user messages.

In addition to pulling friends to build groups, WeChat has also implemented the ability to build groups face-to-face.

4. Face-to-face group building

The user initiates a face-to-face group creation and inputs a 4-digit random code, and users around the group can join the group chat after inputting the random code. The face-to-face group creation function usually involves the following data table design and core business interaction processes.

4.1 Database table design

1. User table: store user information, including user ID, nickname, avatar, etc. 2. Group table: store user information, including user ID, nickname, avatar, etc. 3.
2. Group table: store group information, including group ID, group name, creator ID, number of group members, etc. 3. GroupMember table: store group information, including group ID, group name, creator ID, number of group members, etc. 4.
3. GroupMember table: associated with user and group, including user ID and group ID. 4. RandomCode table: associated with user and group, including user ID and group ID. 5.
4. RandomCode table: stores the random code for face-to-face group creation and the associated group ID.

4.2 Core Business Interaction Flow

User A initiates a face-to-face group in the mobile application, enters a random code, passes the verification, and waits for users in the surrounding area (within 50 meters) to join. At this time, the system stores the user information in the cache as a HashMap and sets the expiration time to 3min.

1

{random code, user list [User A (ID, name, avatar)]}

User B initiates a face-to-face group creation on another cell phone, inputs the specified random code, if there is such a random code around the user, the user enters the same group chat waiting page, and can see the avatars and nicknames of other group members.

At this point, in addition to obtaining all user information based on the random code, the system will also update the user information in the cache in real time.

When the first user clicks Enter the group to join the group chat, the system stores the generated random code in the RandomCode table and associates it with the newly created group ID to update the number of group members.

Then, the system stores the user information and the newly generated group chat information in the Group、GroupMember table

member joins and refreshes the group member information

Later, when user B and C join the group chat with the random code, the mobile client sends a request to the server backend to verify whether the random code is valid. The server backend verifies the random code and checks whether the random code exists in the cache and whether it is within the validity period.

Then, it determines whether the current group members are full or not (at present, the maximum number of group chats created by ordinary users is 500). If the validation passes, the server back-end adds users B and C to the group member table GroupMember and returns a successful response.

When the mobile client application receives the success response, it updates the list of group chats for users B and C to show the new group chats they have joined.

Other Technical Components

In this way, user A successfully creates a face-to-face group by creating a random code and scanning the QR code by the surrounding users. This feature involves several technical components, including distributed caching, database, QR code generation and validation.

Meanwhile, a quite important capability in the face-to-face group building process is to identify the user’s area, e.g., within 50 meters. This can be done by using Redis’ GeoHash algorithm to get information about all users within a range.

Due to space limitations, here does not expand the details, and want to understand more and QR code generation and location algorithm of the details, you can see my previous article: [I heard that you will be architectural design? Come on, get a bus & subway ride system].

5. Message sending and receiving

When a member speaks in a WeChat group, the system needs to ** handle the distribution of the message, notify other members, and ensure that the message is displayed**. The following are the detailed interaction steps for this function, as well as the database storage scheme.

5.1 Interaction Flow

The message sending and receiving timing diagram is shown below:

1. User A sends a message with an image, video or audio to the group.
2. The mobile client application uploads the message content and media files to the server backend. 3.
3. The server backend receives the message and the media file, stores the message content in the Message table, and stores the media file in the distributed file storage cluster. **In the Message table, not only the MediaID of the media file is recorded to associate the message with the media, but also the thumbnail image, video cover image, and so on.
4. The server backend broadcasts the message to all cluster members. The mobile client application receives the message and loads the corresponding presentation based on the message type (text, image, video, audio).
5. when the user clicks to view the image, video or audio thumbnail, the client application fetches the corresponding media file path from the object storage cluster based on the MediaID and displays it to the user.

This process ensures that messages and media files are stored and displayed efficiently. Users can upload and view various types of media data, while the server backend realizes effective message storage and presentation by associating Message with information in the object storage server.

5.2 Message Storage and Presentation

Saving and displaying user’s image, video or audio data in WeChat Groups usually requires the design of data storage and display. In addition to the user table and group table mentioned in the face-to-face group building function above, the following table structures are needed:

1. Message table: Used to store messages, each message has a unique MessageID, message type (text, picture, video, audio), message content (text, picture thumbnail, video cover image, etc.), sender UserID, receiver group GroupID, send time and other fields.
2. Media table: stores media data such as pictures, videos, audios uploaded by users. Each media file has a unique MediaID, file path, uploader UserID, upload time, and other fields. 3.
3. MessageState table: Used to store user’s message state, including MessageID, UserID, whether it is read or not. When the message is pushed, the unread count is calculated through this table, and uniformly pushed to the user, and a small number representing the unread count of the message is displayed on the offline user’s phone.

As we know, MySQL triggers a full table scan every time you query a select count type statement, so loading the unread count of messages is slow every time.

For query performance, we can store the user’s message count into Redis and record an unread value in real time. And, when the unread count is greater than 99, the unread count is set to 100 and will not be increased.

When pushing user messages, ** as long as the unread count is 100, set the number of pushed messages to 99+ to improve storage performance and interaction efficiency. **

6. Grab Red Packet

Grab Red Envelope allows users to send any number and amount of red envelopes in a group chat, and group members can grab red envelopes with random amounts, but it is necessary to ** ensure that the amount of red envelopes for each user is not less than $0.01**.

The detailed interaction flow of grabbing red packets is as follows:

1. the user receives the notification of grabbing the red packet, and clicks on the notification to open the group chat page.
2. The user clicks on the red packet, and the background service verifies the user’s eligibility to ensure that the user has not yet received the red packet.
3. If the user’s eligibility is verified, the background service allocates the red packet amount and stores the collection record.
4. The user sees the red packet amount in the WeChat group, and the red packet status is updated to “Claimed”.
5. Asynchronously call the payment interface to update the red packet amount to the wallet.

The red packet function requires attention to the database design, real-time red packet grabbing and red packet distribution algorithm.

6.1 Database Design

The fields of the redpack table redpack are as follows:

• **id: ** primary key, redpack id
• **totalAmount: ** total amount
• surplusAmount：* remaining amount
• **total: ** total number of red packets
• surplusTotal：* Remaining red packets total amount
• userId：* The ID of the user who sent the red packets.

This table is used to record how many red packets a user has sent and the remaining amount to be maintained.

The redpack record table redpack_record is as follows:

• id: primary key, record ID
• **redpackId: ** redpackId, foreign key
• **userId: ** userId
• **amount: ** amount grabbed

The record table is used to store information about the specific redpacks grabbed by users, and is also a sub-table of the redpack table.

6.2 Real-time

1. Send red packet

1. After the user sets the total amount and number of red packets, add a data in the red packet table and start sending red packets.
2. In order to ensure the real-time performance and the efficiency of grabbing red envelopes, add a record in Redis, .
3. Grab the red envelope message pushed to all group members

2. Grab Red Envelope

Since 2015, WeChat Red Envelope’s grabbing red envelopes and splitting red envelopes have been separated, and users need to perform two operations after clicking on grabbing red envelopes. That’s why you can sometimes grab a red packet, but when you click on it, you’ll find that the red packet has already been claimed**.

The interaction steps of grabbing red envelopes are as follows. 1:

1. Grab the red packets: the grabbing operation is done in the Redis cache layer, ** through the atomic decrement operation to update the number of red packets **, after reaching 0, it means that all the red packets have been grabbed.
2. opening red packets: when opening red packets, the first real-time calculation of the amount of money, generally through the two times the average method (i.e., between 0.01 and 2 times the remaining average). 3. red packet records: the user to get the red packet records, the user to get the red packet record.
3. Red Packet Records: After the user obtains the amount of the red packet, the number and amount that have been received are totaled through the transaction operation of the database, and the red packet table and record table are updated.
4. transfer: in order to improve efficiency, the final transfer is an asynchronous operation, which is why during the Chinese New Year, the red packet cannot be seen in the balance immediately after receiving it.

6.3 Red Packet Allocation Algorithm

When the red packet amount is allocated, there are two implementation options: real-time splitting and pre-generation because it is randomly allocated.

1. Real-time splitting

Real-time splitting refers to the process of real-time calculation of the amount of each red packet when ** grabbing red packets, in order to realize the process of red packet splitting.

This requires us to design a good splitting algorithm, so that the red packet splitting has been to ensure that the amount of the subsequent red packet to be split can not be empty.

When splitting in real time, it is not easy to do the split red packet amount obeys the normal distribution law.

2. Pre-generation

Pre-generation, refers to the red packets ** before ** have completed the red packet ** amount of split **, grab the red packet is only taken out in order to split the amount of red packets.

This way of splitting algorithm requirements are lower, you can split the red packet amount of randomness is very good, but usually need to be combined with the use of queues, and the need to design a table to store the split amount of red packets.

3. Twice the mean method

Considering the advantages and disadvantages of the above, as well as the number of people in the WeChat group chat is not large (currently up to 500 people), so we use real-time splitting, using Double Mean Method to generate random red envelopes, only to meet the random can be, do not need to be normally distributed.

Therefore, there may be a large difference between the red packets, but it is more exciting, isn’t it 🐶.

Random numbers generated using the doubled mean method will have a random amount between 0.01 ~ 2 each time.

Assuming that the remaining amount of the current red packet is $10 and the number of remaining red packets is 5, 10/5 = 2, then the amount of red packets that the current user can grab is: 0.01 ~ 4 yuan.

4. Algorithm Optimization

Although the random red packets generated by the twofold mean method are close to the average, I have previously seen a similar statement on a forum: The randomness of the WeChat red packet amount has a relationship with the timing of receiving it, especially if the amount is not high.

So, Xiao ❤ spent a huge amount of money to send multiple red envelopes in the WeChat group, and came to the conclusion that after sending 4 red envelopes totaling 0.05, the amount of red envelopes received by the last person must be 0.02.

No exceptions:

So, the probability is that the red packet amount algorithm is not randomly assigned, but has been processed before handing out the red packets. For example, before the red packet amount is generated, it is first generated into a non-existent red packet, which totals

And when the red packet amount is assigned, 0.01 is added to the random value base of each red packet as a way to ensure that the minimum value of each red packet is not 0.

So, suppose the user sends out 3 red envelopes with a total amount of 0.04, he needs to first extract 3*0.01 to the “fourth” non-existent red envelope, so the random value of the red envelope grabbed by the first person is 0 ~ (0.04-3*0.01)/3.

The quotient of the divisor is taken down to two decimals, 0 ~ (0.04-3*0.01)/3 ==> (0 ~ 0) = 0, plus the previously extracted guaranteed value of 0.01, because of the concern about red packet overruns, so the first two grabbed red packet amounts are both 0.01. The amount of the last red packet is the red packet balance, i.e. 0.02.

The algorithm logic is implemented in Go as follows:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62

import (
   "fmt"
   "math"
   "math/rand"
   "strconv"
)

type RedPack struct {

    SurplusAmount float64 

    SurplusTotal int 

}

func remainTwoDecimal(num float64) float64 {

    numStr := strconv.FormatFloat(num, 'f', 2, 64)

    num, _ = strconv.ParseFloat(numStr, 64)
    return num
}

func getRandomRedPack(rp *RedPack) float64 {
    if rp.SurplusTotal 0 {

        return 0
    }

    if rp.SurplusTotal == 1 {
        return remainTwoDecimal(rp.SurplusAmount + 0.01)
    }

    avgAmount := math.Floor(100*(rp.SurplusAmount/float64(rp.SurplusTotal))) / float64(100)
    avgAmount = remainTwoDecimal(avgAmount)

    rand.NewSource(time.Now().UnixNano())

    var max float64
    if avgAmount > 0 {
        max = 2*avgAmount - 0.01
    } else {
        max = 0
    }
    money := remainTwoDecimal(rand.Float64()*(max) + 0.01)

    rp.SurplusTotal -= 1
    rp.SurplusAmount = remainTwoDecimal(rp.SurplusAmount + 0.01 - money)

    return money
}

func main() {
    rp := &RedPack{
        SurplusAmount: 0.06,
        SurplusTotal:  5,

    }

    rp.SurplusAmount -= 0.01 * float64(rp.SurplusTotal)
    total := rp.SurplusTotal
    for i := 0; i 

Print results:

0.01、0.01、0.01、0.01、0.02

As expected!

7. Summary

Behind WeChat’s group chat and red packet grabbing features lie complex interaction techniques and well-designed product experiences. Through these core components, database tables, and detailed interaction processes, users are able to easily participate and enjoy the convenience of the group chat system.

And, the addition of these fun-filled features is one of the reasons why WeChat has so many users, right?

The system design of WeChat’s group building feature is not just a display of technological magnificence, it’s part of the magic of digital socialization.

1. Introduction
2. Three Link Generation Methods
3. The Redirection Process
4. Cache Optimization
5. Designing for High Availability
6. Postscript

**01 Introduction

**1) Background

This is a system design question in my interview with the department of “Byte Shake”, the position is “back-end senior development engineer”, asked in the second interview. At the beginning, the interviewer smiled and asked me to introduce myself, and then talked about the project.

After talking about the project flawlessly and writing an algorithm question, the interviewer started to ask questions.

The interviewer began to ask: young man, resume inside wrote familiar with architectural design is not it, then you know program design ‘three high’ refers to what?

I thought to myself, that is not due to the programmer’s system is not reliable, the leadership is not the right person, every day overtime to change the BUG, resulting in a young age are high blood fat, high blood pressure and high blood sugar!

However, since this is an interview, the leader is certainly not willing to listen to this, so I answered: program three high, that is, the system design needs to consider the high concurrency, high performance and high availability:

• High concurrency is in the process of system development, the need to ensure that the system can simultaneously process many requests in parallel;
• High performance means that the program needs to take up as little memory and CPU as possible and process requests as fast as possible;
• High Availability usually describes that the system is not serviceable for a short period of time, for example, no more than 31.5 seconds of downtime throughout the year, commonly known as 6 9s, which means that availability is guaranteed for 99.9999% of the time.

So, the interviewer nodded slightly, thinking that the young man is not bad, since this can not be difficult for you, then I have to come up with a big trick, let’s have a system design questions!

**2) Statement of Requirements

As we all know, when business scenarios need to send the user a network address or QR code, due to the length of the address is relatively long, usually in order to take up fewer resources and improve the user experience. For example, the address of the Google search term “computer” is as follows:

1

https://www.google.com/search?q=%E8%AE%A1%E7%AE%97%E6%9C%BA&ei=KNZ5Y7y4MpiW-AaI4LSACw&ved=0ahUKEwi87MGgnbz7AhUYC94KHQgwDbAQ4dUDCBA&uact=5&oq=%E8%AE%A1%E7%AE%97%E6%9C%BA&gs_lcp=Cgxnd3Mtd2l6LXNlcnAQAzIECAAQQzIFCAAQgAQyBQgAEIAEMgUIABCABDIFCC4QgAQyBQgAEIAEMgUIABCABDIFCAAQgAQyBQgAEIAEMgUIABCABDoKCAAQRxDWBBCwAzoLCC4QgAQQxwEQ0QM6FggAEOoCELQCEIoDELcDENQDEOUCGAE6BwguENQCEENKBAhBGABKBAhGGABQpBZYzSVglydoA3ABeACAAZ0DiAGdD5IBCTAuNy4xLjAuMZgBAKABAbABCsgBCsABAdoBBAgBGAc&sclient=gws-wiz-serp

Obviously, it is not “decent” to send this long list of URLs to users. Moreover, when encountering some systems with word limits, such as microblogging posts with word limits, it is definitely not possible to send such a long link address. In general, most of the SMS links are short links:

So, in order to improve the user experience, as well as the daily business needs. We need to design a short link generation system, in addition to the realization of business functions, we have to serve the national network address. With such a large number of users, how should the data be stored and how should high concurrency be handled?

02 Three methods of link generation

1) Requirements Analysis

I thought to myself, this interviewer looks “kind eyes” and smiling, but the topic is not simple, this type of system needs to be considered too many points, absolutely can not be taken lightly.

So, I was from the link generation, URL access, cache optimization and high availability of four aspects of the design.

First of all, the generation of short link address, you can consider using UUID or self-incrementing ID. for each long link to short link address, must generate a globally unique short link value, otherwise there will be a conflict. So, short links are characterized by:

• The amount of data storage is large, the national URL is at least a million short link addresses need to be generated every day;
• Concurrency is not small, encountered at the same time to access the system, according to 3600 seconds a day to calculate, the average number of requests per second at least thousands;
• short links can not be repeated, otherwise it will cause data access conflicts.

**2) Snowflake Algorithm

First of all, to generate short links, you can use the snowflake algorithm + hash to achieve.
The snowflake algorithm is a unique number generated in a distributed scenario based on timestamps, different machine IDs, and sequence numbers. Its advantage is that it is simple and easy to use as you go.

Through the snowflake algorithm to get the unique number, and then hash mapping, the number will be converted to a random string, if the short chain string is longer, you can directly take the first 6 bits. However, because the result of hash mapping may conflict, so the hash algorithm is more demanding.

2) 62 Progressive Number Generation of Short Links

In addition to the snowflake algorithm, you can also use the number of 62 (A-Za-z0-9) to generate short link addresses. First get a self-incrementing ID, and then convert this value to a string of hexadecimal (a-zA-Z0-9), a billion digits converted to just five or six (100 million -> zAL6e).

The short link server domain name, and this string for splicing, you can get the URL of the short link, such as: t.cn/zAL6e.

And generating the self-incrementing ID needs to consider the performance impact and concurrency security, so we can make an issuer through Redis’ incr command, which is an atomic operation, so we don’t have to worry about the security of the numbers. And Redis is an in-memory operation, so it is quite efficient.

3) Random number + Bloom filter

In addition to incrementing the ID, we can also generate random numbers and then convert to hexadecimal to generate short links. However, since the random numbers may be duplicated, we need to use Bloom filters to de-duplicate them.

The Bloom filter is a cleverly designed data structure that works by hashing a value multiple times, mapping it to different bit bits and recording it. When the new value is used, the same hash function is used to compare whether there is a value on each bit bit: if there is no value on any of these bit bits, the number is unique; otherwise, it may not be unique.

Of course, this may produce a misjudgment, ** Bloom filter can certainly find duplicate values, but ** ** may also be non-duplicated values judged as duplicates **, the misjudgment rate of roughly 0.05%, is an acceptable range, and the Bloom filter is extremely efficient.

Therefore, through the Bloom filter, we can determine whether the generated random number is duplicated or not: if it is duplicated, a new one will be generated; if it is not duplicated, it will be deposited into the Bloom filter and the database, thus guaranteeing that the random number fetched each time is unique.

**4) Store the short link to the database **

When storing the database, different databases may be chosen because of the amount of inventory and technology stack. However, since MySQL is used more in the company department and the current topic does not mention technology selection, we choose MySQL as the persistent database.

Whenever a short link is generated, the mapping relationship from short link to long link needs to be stored in MySQL with a unique index, i.e. zAL6e -> real URL.

• 03 Redirection process

  When the browser accesses the short link service, it fetches the original URL based on the short link address and then redirects the URL. We usually have two redirection methods:

  • One is to return a 301 response code to the browser for permanent redirection, so that it can directly access the real URL address in the future;
  • One is a 302 temporary redirection, so that the browser currently this time to visit the real URL, but subsequent requests are still based on the short chain address access.

  Although the 301 browser only one request, the follow-up can be directly from the browser to get a long link, this method can improve access speed, but it can not count the number of visits to the short link.

  So according to the business needs, we usually choose 302 redirection.

  04 Cache Design

  Since the short link is distributed to multiple users, it may be accessed multiple times in a short period of time, so the long link can be put into redis cache after writing/fetching it from MySQL.

  1) Add to cache

  And, the correspondence between short and long links is usually not modified frequently, so the consistency between the database and the cache is ensured by a simple Bypass Cache Mode:

  • When reading (Read) data, if the cache is not hit, the DB is read first, and the data is taken out of the DB and put into the cache, while the response is returned;
  • When writing (Write) data, first update the DB, then delete the cache.

  When the user needs to generate a short link, first go to this mapping table to see if there is a corresponding short link address. If there is one, return it directly and increase the expiration time of the key-value by one hour; if not, regenerate it and store the corresponding relationship in the mapping table.

  The cache elimination strategy can be chosen:

  • LRU: Least Recently Used, the short-link address that has been read and written recently as hot data can always exist in the cache, eliminating those short-link keys that have not been accessed for a long time;
  • LFU: Least Frequently Used, the recent least frequently used algorithm, the recent short chain address with high access frequency can be used as hot data, eliminating those short chain keys with lower access frequency.
  2. Cache penetration

  However, the use of cache can not prevent some anomalies, such as “cache penetration”. The so-called cache penetration is to query a cache and database do not exist in the short link, if the concurrency is very large, it will lead to all the cache does not exist in the request to hit the MySQL server, resulting in the server can not handle so many requests and blocked, or even crash.

  Therefore, in order to prevent the miscreants from attacking the server in a similar way as “cache-penetration”, we can adopt two methods to deal with it:

  • Cache the non-existing short link address, the key is the short link address, the value value is empty, the expiration time can be set to a shorter time;
  • Bloom filter will have a short link hash many times after the storage, when there is a short link request, first through the Bloom filter to determine whether the address exists in the database; if not, it means that the database does not exist in the address, it will be returned directly.

05 High Availability Design

Since caching and database persistence rely on Redis and MySQL, high availability of MySQL and Redis must be guaranteed.

1. MySQL High Availability

The MySQL database uses master-slave replication to separate reads and writes; the master node is used for writes and the slave node is used for reads, and it can be keptalived to achieve high availability.

The principle of Keepalived is to use virtual IP to detect multiple nodes in the portal, select a hot standby server as the master server, and assign it a virtual IP through which external requests can access the database.

At the same time, Keepalived detects the availability status of multiple nodes in real time, and when a server is found to be down or faulty, it will be kicked out of the cluster. If this server is the master, keepalived will trigger an election operation to elect another server from the server cluster to be the master and assign it the same virtual IP to complete the failover.

And, with Keepalived’s support, these operations don’t require human involvement, just the repair of the failed machine.

2. Redis High Availability

Because of the high concurrency of big data scenarios, write requests all fall on the master node of Redis, the pressure is too great. If you insist on increasing memory and CPUs in this vertical expansion, then a machine facing the gradual increase in disk IO, network and other pressures, which will also affect performance.

So Redis uses cluster mode to realize data sharding. Moreover, a sentinel mechanism is added to ensure the high availability of the cluster. Its basic principle is that the sentinel nodes monitor all the master and slave nodes in the cluster, when the master node is down or after a failure, the sentinel nodes will mark it as subjective offline; when enough sentinel nodes will mark the Redis master node as subjective offline, it will be changed to the state of the Objective offline.

At this point, the sentinel nodes elect a lead sentinel through an election mechanism to perform a failover operation on the Redis master node to guarantee the high availability of the Redis cluster, and this entire process requires no human intervention.

3. System Fault Tolerance

Before the service goes live, it needs to do a full assessment of the business volume, as well as performance testing. Do a good job of limiting the flow, fusion and service degradation logic, for example: the use of token bucket algorithm to achieve the limitation of the flow, hystrix framework to do the fusion, and will be commonly used configurations into the configuration center that can be hot update, so as to facilitate the pair of its real-time changes.

When the business volume is too large, the synchronous task is changed to asynchronous task processing. Through these service governance solutions to make the system more stable.

06 Postscript

When I finished the last word, the interviewer looked at me with admiration and doubt in his eyes. I thought that he should have been impressed by my performance, and that the interview should have been a sure thing.

However, surprisingly, the interviewer did not comment on the architectural design, but only looked at me and said, “That’s it for today’s interview, do you have anything you want to ask?”

Now, it was my turn to be shocked, so did I pass or not? But to give a sentence ah, so I asked, “through this interview, you think I have what aspects need to be improved?”

“Algorithms and projects need to be practiced more, but I found a good thing about you.” The interviewer laughed and then said, “You’re pretty good at memorizing eight-legged essays!”

My hanging heart finally dropped, and I thought to myself, “Oh, that’s steady~”

1. Introduction
2. Requirements Analysis
3. Outline Design
4. Detailed Design
5. Publish/Subscribe Issues
6. Postscript

1. Introduction

When I interviewed for the “Baidu.com” department’s “Backend Senior Development Engineer” a year ago, I was quite impressed by one of the interview questions during the third interview.

I remember that the interviewer was very professional, asked a lot of project-related and thoughtful questions, and a few of my usual development process in the accumulation of high-quality BUG, also asked.

I was complacent, thinking that the test will be, answered all right, the interviewer began to make things difficult: “Since you think quite a lot on the project, then I will test you on a project design topic! First of all, these questions are not meant to be difficult for anyone, but just to see the depth and breadth of the candidate’s technical knowledge!”

Translated with www.DeepL.com/Translator (free version)

“Ah, right, right, right! It’s like this”, although the heart of a hundred apprehensions, it is true that the daily requirements are CRUD (add, delete, change, and check), the architectural design has not been done in the project for a long time, but still showed no panic, the color of a ready mind.

After all, the interview book said: half of the time in the interview are playing psychological warfare! As long as you do your homework, you can defend yourself confidently in the examination room, and we won’t be defeated by the interviewer’s one or two hour attack.

2. Requirements Analysis

The interviewer asked, “As a popular social APP, do you usually use Weibo? Can you tell me a few common functions of microblogging?”

Microblogging has not really been used much, but I just recently read a system design plan about microblogging, so I answered without panic: “The common functions of microblogging are brushing microblogging, posting and user attention, in addition, users can also like, comment, favorite and forward microblogging.”

“Good, then if you are now allowed to design a microblogging system, combined with the several core functions you just said, how would you design it? The system needs to be considered for high concurrency, high performance, and high availability.”

After getting the “Product Requirements”, I first built a use case diagram of the core functions in my mind as follows:

Translated with www.DeepL.com/Translator (free version)

Each requirement is described as follows:

• Brush microblogging: users open the microblogging homepage on the mobile APP side, displaying the most recent microblogs published by the friends they follow, sorted by the most recent time;
• Tweeting: Users can post content up to 140 characters of text, which can contain pictures and videos;
• Follow friends: users can follow other users, and the followers can see the information and number of followers.

3. Outline Design

The business functions of Weibo are not difficult to understand, but ** the concurrency and data volume** are very large:

• 1 billion level of user volume, on average thousands of post counts per user, each user can follow thousands of friends;
• high concurrency, the average page access of 100,000 level per second, the posting volume of 10,000 level per second;
• Uneven distribution of users, the number of posts or the number of fans of some star users exceeds that of ordinary users by several orders of magnitude;
• Uneven distribution of time, a user may suddenly become a hot user at a certain point in time, and its fans may also steeply increase by several orders of magnitude.

It has the characteristics of a typical social system, which can be summarized into three points: massive data, high access, non-uniformity, and then from its most common follow friends, brush microblogging, posting three functions to do the outline design.

3.1 Follow Friends

Translated with www.DeepL.com/Translator (free version)

If you open the main page of a blogger in Weibo, the page contains the most basic features: Ta’s followers, Ta’s fans, and we can click to follow this blogger. From the blogger’s main page, we can also access his attention and fans sub-pages:

• Attention page, which displays information about all the users that the user follows.
• Follower page, which shows all the fans of the user.

In the above page, the user can follow a certain user, and can also delete fans, that is, cancel the attention of a certain other user, the business interaction of the function is as follows:

When a user follows a friend, it needs to go through a load balancing server first, and then send the request to Weibo’s user server cluster, where it involves displaying the user’s information, so the user servers may access object storage servers for images and videos, as well as take out text data from Redis or MySQL.

Finally, if the user’s follow status is modified, the information is written to the Redis and MySQL clusters.

3.2 Brush Microblogging

The core of the microblogging system is to solve the problem of high concurrency. The overall deployment model of the system is as follows:

First, we use CDN (Content delivery network) to quickly return user requests, which is based on the principle of deploying servers in regional nodes that are widely used by users, and when a user accesses the system, the request is distributed to the nearest servers through the Global Load Technique, and they provide services directly to the user:

When CDN is not used, each user request will directly reach the system’s application server cluster, which will generate great traffic pressure on the system’s servers; when CDN is used, the user requests will be distributed to the nodes in the user’s neighborhood according to the target servers returned by the CDN load balancer.

The advantage of using CDN is that it can greatly avoid network congestion and make content transmission faster and more stable. CDN can be regarded as a system cache, which is especially suitable for microblogging scenarios where data is seldom changed, and under normal circumstances, CDN can filter out more than 90% of the requests and return data directly.

Under normal circumstances, CDNs can filter out more than 90% of the requests and return the data directly. Therefore, when a user accesses the Weibo system through a CDN, the vast majority of the requests will be hit by the CDN cache. In other words, more than 90% of bandwidth-consuming requests such as images and videos can be digested by the CDN. Requests that are not hit by the CDN will arrive at the reverse proxy server in the data center, which checks whether the local cache has the content needed for the request. If there is, it will return directly; if not, it will go to the distributed object storage cluster to get the relevant images and videos, or get the microblogging text content from the application server application.

When fetching content from the microblogging application server cluster, it will first retrieve the latest tweets posted by the current user’s friends from the Redis cache server and build a result page to return. If the number of tweets cached in Redis is less than 20, it will continue to look up the data from the MySQL database.

4.3 Writing Tweets

Instead of CDNs and reverse proxies, tweets are written directly to the application server cluster via load-balanced servers. The application server writes the published tweets to the Redis cache cluster on one hand, and to the MySQL sharded database on the other hand.

Note that when we write to the database, if we write to the database directly, when a highly concurrent write request suddenly arrives, it may cause the database to overload, which may lead to system blocking or crashing (refer to the “Cache Avalanche” problem). Therefore, database write operations can be written to a message queue (e.g., Kafka cluster), and the message queue’s consumer program will consume messages from the message queue at a certain rate and write them to the database to ensure that the DB’s load pressure does not surge and cause abnormalities.

4. Detailed Design

4.1 Table Design

User, fan, follow and post tables:

• user table: primary key id, user information (name, avatar, registration time, v-authentication, cell phone number, etc.)
• relation table: primary key id, followId, attentionId [fan and follower IDs]
• post table: primary key id, userId, postTime[posting time (accurate to milliseconds will do)], content

Index Optimization

post table can use combination index userId+postTime to query the recent posts of a user, here the combination index is used as a secondary index, so it needs to go back to the table. In order to reduce the number of times to go back to the table, we can splice the userId and timestamp bit postId as the primary key: for example, the first 20 bits as the timestamp is accurate to milliseconds + the userId, converted into a hexadecimal (0-9a-zA-Z) string as the postId, which also ensures that the primary key is incremental.

However, this will cause the index tree to take up more space, and the query is not as fast as a pure numeric primary key, so you can finally choose the most suitable primary key type according to the actual situation and compare the advantages and disadvantages of the two.

4.2 Splitting a library into tables

When the data volume of follow, attention, and user tables exceeds 10 million or 100 million, the pressure of microblogging to read and write the database is very large, which is definitely not bearable for a single database. Therefore, the microblog DB needs to be a distributed database with sharded deployment.

Vertical Table Splitting

The above relation table [id, followId, attentionId] stores the information of followers and attention, when the number of users increases, there will be a problem, that is, it is not easy to choose the key for splitting the database and table:

• If followId is chosen as the hashKey, the query for the current user’s attention list will be on the same slice, but the query for all of the user’s followers will need to be on multiple slices;
• If we choose attentionId as hashKey, we will be on the same slice when querying all followers of a user, but we need to be on multiple slices when querying all followers of the current user.

So we split the relation table into:

• follow table: primary key id, userId, followId
• attention table: primary key id, userId, attentionId

Horizontal Splitting

Horizontal splitting means deploying a distributed database using hash slicing, the slicing rules can be user ID or post ID.

If you slice by user ID, all the posts made by the same user will be saved to the same database server. The advantage is that when the system needs to find out the tweets made by a certain user, you only need to access one server to accomplish it. The disadvantage is that for a star user, there will be a lot of data access, and the access of hot data leads to excessive load pressure on that server. Similarly, if a user posts tweets frequently, this can lead to excessive data growth on a single server.

If you slice by post ID, although you can avoid the hotspot aggregation problem caused by user ID slicing, when looking up all the microblogs of a user, you need to access a randomly sliced database server, which puts too much pressure on the entire database server cluster.

All things considered, the hotspot problem caused by user ID sharding can be improved by optimizing caching. The problem of frequent tweets by a certain user

4.3 Hot Users Problems

After splitting the table, although the problem of fast table checking has been solved to some extent, the query for some hot star users still needs to be optimized. For example, the following scenarios:

• The hot star user has a lot of fans, and the number of rows scanned when querying the number of fans by following the table to query count is very large, and this inefficient operation will be expanded due to the increase in the number of fans.
• When brushing microblogs, star users have many posts that are repeatedly viewed, if they go to the DB to get them every time a fan goes to query, there will undoubtedly be serious performance pressure.

To solve the DB query problem in the above two scenarios, caching can be introduced. But the cache space is limited, we necessarily can not cache all the data, set a good cache elimination strategy is the focus of our discussion.

Time Elimination Strategy

For hot topics, both hot posts and hot users need to be added to the cache, and the cache elimination strategy can be set as a time elimination algorithm.

All the tweets published in the last n days are cached, and when a user refreshes the tweets, he/she only needs to find them in the cache. If a user gets a list of 10 tweets, it will be returned directly to the user; if the number of tweets in the cache is not enough, it will be looked up in the database.

So, we can cache all the tweets published within 7 days, where the cache key is the user ID and the value is the list of post IDs published in the last 7 days. At the same time, the post ID and post content are also cached as key and value respectively.

Local Cache Mode

In addition, for especially popular microblogs, such as celebrities getting married/divorced/growing up in a GUI, high concurrent accesses are concentrated on a single key, which puts a great load pressure on a single redis server. Therefore, microblogging system can enable local cache mode, i.e., the application server will cache popular microblogs in the server memory, so that when users brush microblogs, it will prioritize checking whether the microblogs corresponding to the post IDs are in the local cache or not.

For big V users with more than 500w followers, we can cache all their tweets published within 48 hours to further reduce the pressure of checking hot data.

5. Microblog Publishing/Subscribing Issues

The publish/subscribe problem is the core business problem of Weibo, which is how to quickly get the latest content of all friends after following them.

5.1 Push Mode

When a user publishes a post, he/she immediately pushes the message to his/her fans. However, at this time, the fans are not necessarily online, so the data needs to be stored. In this way, every time a user adds a new post, the user needs to push the post to the slice of the DB where its followers are located, and the followers can directly query the data stored in their own slice every time they browse the new message.

An obvious problem with the push model: ** If a user has tens of millions of followers, then every time the user posts a tweet, he needs to insert tens of millions of records into the subscription table, i.e., “write proliferation “** . There is no lack of zombie fans (users who are online very infrequently) among the followers, which brings the consequence that the database is under very high pressure, leading to blocking or crashing.

5.2 Pull Mode

When a user publishes a post, it is only saved in his/her own business table. When the followers come online, they go to the followers table to read the posts and return them in chronological order.

The problem of pull mode is: if a user follows 500 star users, each time the query needs to go to each slice to query the posts published by different stars, ** a star has tens of millions or even billions of fans, which means that there may be tens of millions or even billions of read data operations at the same time, i.e., “read diffusion “** , the problem is that the database slice of the read data decompression is not reflected in the effect.

Therefore, the microblogging system firstly needs to limit the number of users’ attention, the microblogging ordinary users’ attention limit is 2,000 people, and the VIP users’ limit is 5,000 people. Secondly, minimize the number of database queries when refreshing the microblog page, and use the cache to read more posts.

5.3 Combination of Push and Pull

We found that even if we limit the number of friends for microblogging users, it is difficult to solve the subscription/publishing problem of the microblogging system with a single “push mode” or “pull mode”, so we finally adopt the “combination of push and pull” mode, which is realized in the following two ways.

1) Distinguish between big v-stars

For big v-star users (with more than 500w followers), in order to prevent write proliferation, we only need to synchronize the data to 100 database slices (assuming there are 100 data slices), which need at least three fields: userId, postId, postTime. no matter how many followers there are, there will be only 100 replicas, which avoids many zombie followers who will never go online due to data write. This avoids a lot of zombie fans that will almost never be online because of the data writing waste of network performance and storage resources.

For ordinary users, or continue to use the push mode, so that most users in the read the latest posts only need to read the corresponding slice of their own users can get the data.

This is a simpler way to design, but when the user is brushing the microblogging, ** the query process may need to query twice, respectively, the post information under the table of their own subscription, and the post information under the table of the concerned user’s publication **.

2) Distinguish online status

The second way is to judge push and pull according to the user’s state: if the user is currently online, push mode is used, and the system will create a list of friends’ latest published tweets for them in the Redis cache. If a friend who follows has a new tweet published, the post information is immediately inserted into the head of the list, and when that user refreshes the tweet, he/she just needs to return this list.

If the user is not online, the system deletes the list, and when the user logs in again to refresh, the list is rebuilt for them using pull mode.

How to determine the user online? On the one hand, it can be determined by the time interval of user operation, i.e. heartbeat mechanism; on the other hand, it can also be predicted the online time of the user by machine learning, and utilize the idle time of the system to build the latest microblog list for them in advance.

6. Postscript

The interviewer listened to my analysis, thinking that this kid is thinking quite comprehensively, but the face can not show it. So he nodded slightly and said, “That’s all the questions I have, so is there anything you want to ask?”

So with the idea of hijacking the interview, after asking a few painless business questions, the interview was over. After all, the interviewer is going to catch me asking about the details of the architectural design, and I may never remember it again.

I’m not sure if I’m going to be able to remember the details of the architectural design.

I not only sigh this eight-legged text, architecture questions, ** memorize is really slow, forget is really fast ** ah!

Fortunately, usually have the habit of summarizing the architectural design, not to make a fool of myself, and then will also put all the architectural design encountered during the interview process on the inside of the personal GZH, those who need to move to take their own Ha~!

1.1 Reservoir flooding

Shenzhen had just sent away Typhoon Sura and welcomed Anemone.

On the same day, schools in Shenzhen once again announced the closure of classes.

A student feedback: in shenzhen school is too “tough”, the opening week of school, closed twice.

Workers inside OS: you guys just steal the joy, after all, as long as the traffic does not stop, we are not going to stop working.

Shenzhen Transportation: The reservoir is full, what’s the matter with my transportation? After all, our high availability design is not a cover.

1.2 Availability of Shenzhen transportation system

If the Shenzhen traffic as an application system, from the point of view of software development, flooding is an unexpected attack, the Shenzhen reservoir flooding event is actually ** to maintain the stability of the system to carry out bulk data deletion **.

Although the instantaneous data is very large, resulting in a short period of unavailability of Shenzhen’s transportation system. However, through the operation of the drainage system, temporary diversion and other measures, Shenzhen transportation, a system that serves more than 10 million users, quickly operated again.

I have to say, that’s pretty strong!

Interview Review

What does software usability have in common with transportation usability?

The thought spreads infinitely, reminding me of an interview question from the second interview of Tencent a long time ago.

Interviewer: we know that distributed systems have three high, high performance, high concurrency and high availability. ** taking into account the performance and concurrency of the system, from what aspects will you consider to design a highly available system **?

Combined with the handling of Shenzhen traffic, I will introduce the basic concepts of high availability, performance estimation, system testing, and flow-limiting means respectively.

2. High Availability Design

2.1 What is High Availability

High Availability (HA) of a software system is one of the factors that must be considered in the architectural design of a distributed system, and it is usually measured by the time that a system provides a service.

If we take Shenzhen traffic as a system, assuming that the traffic can always run, the availability of the system is 100%.

Of course, as with rainstorms and floods, the system can suffer from sudden traffic surges, or hacker attacks, so 100% availability of the system is hard to guarantee.

!

For every 100 units of time the system is running, there will be 1 unit of time when the service is not available, the availability of the system is 99%.

Large software systems, such as Taobao, WeChat, etc., pursue a high availability goal of 99.9999%, commonly known as 6 9, i.e., no more than 31.5 seconds of downtime throughout the year.

2.2 Performance Metrics

When we design a system, we need to do a measurement of the system load capacity, i.e. Performance Estimation, and then target the high availability design.

When doing performance estimation, we have to understand a few metrics first.

Response Time

Response time refers to the time it takes for a client to receive the response data from the time it sends out a request, which is the most important performance indicator of the system, directly reflecting the speed of the system to process the request.

Why is the response time is the most important indicator, we take the Shenzhen transportation chestnut, the response time is in the queue time passengers need to wait for a taxi, or in the subway station queuing for the subway time.

!

The length of the response time is a direct reflection of whether the transportation transit status is normal.

Concurrency Count

Concurrency count is the number of requests the system handles at the same time, reflecting the load pressure of the system. When we do performance testing, we usually use multiple threads to simulate the number of concurrent users, each thread acts as a user request, and the number of threads is the number of concurrency obtained in the performance metrics.

The concurrency number of a transportation system can be understood as the sum of the number of passengers in all transportation.

Throughput

Throughput is the number of requests processed by the system per unit of time, which reflects the size of the system’s ability to handle the business. It is generally measured in terms of HPS (Requests Per Second, Hits Per Second), QPS (Queries Per Second, Qeries Per Second) and TPS (Transactions Per Second, Transactions Per Second).

During the normal operation phase of a system, the relationship between these three metrics is: Throughput = Number of Concurrencies / Response Time.

Throughput in a transportation system reflects whether the traffic is flowing smoothly and how much capacity the system is carrying.

2.3 System Testing

After familiarizing ourselves with the performance indicators of the system, we can start to do system testing. The specific process is as follows: ** Continuously increase the concurrency of the system, so as to test out the system’s ability to resist pressure, as well as the system’s performance threshold **.

!

System testing is divided into three phases, including performance testing, load testing and stress testing.

Performance Testing

In the early stage of system design, we will first plan an expected target for performance. During the testing period, continuously apply pressure on the system to verify whether the system achieves the expected performance target within the acceptable range of resources.

Take Shenzhen transportation as an example, we design a subway system at the initial stage, through research and estimation, etc., based on the size of the daily flow of people to design the subway system.

Performance testing can be interpreted as increasing the number of passengers under a computer model to estimate whether the subway system can achieve the expected transportation capacity.

Load Testing

Constantly applying concurrent requests to a system, increasing the stress on the system until one or more of the system’s metrics reaches a safety threshold.

The same is true for transportation systems, if there are too many objects (people or cars) involved in the traffic, the overall transportation capacity will be affected! That’s why all the first-tier cities limit the number of cars on the road.

Stress test

Exceeding the safe load, increasing the number of concurrent requests and continuing to stress the system until the system crashes or no more requests are processed, at which point the number of concurrent requests is the system’s maximum stress tolerance.

Stress test is to measure the maximum capacity of the system bottleneck, take the traffic system as an example, if the number of passengers in a moment to reach the maximum capacity value, Shenzhen traffic will immediately paralyze.

PS：The maximum pressure capability is a theoretical value that may not be reached in practice, but must be estimated during system design.

3. “Flow Limiting” for Highly Available Designs

3.1 Flow limiting

Limit flow in the early stage of system design generally do not need to consider, and when the number of users increases, ** when the system’s processing capacity can not cope with the external sudden increase in traffic access, in order to let the system to maintain stability, must be taken to limit the flow of measures **.

Inside the transportation system, the most common flow restriction is the foreign vehicle restriction.

!

For example, in Shenzhen, foreign cars are only allowed to travel on specified dates (weekends or holidays), or are not allowed to travel during specified peak hours (weekdays from 7am to 9am, and from 5:30pm to 7:30pm), all of which belong to the flow restriction to maintain the normal operation of the transportation system.

1) Traffic Restriction Indicators

TPS

Transactions Per Second, The number of transactions completed per second. This is the most reasonable value to limit the flow, but it is not very practical, because in distributed business systems, transactions often need more than one module to complete.

To limit the flow according to TPS, the time granularity may be very large, and it is difficult to accurately assess the response performance of the system.

In the transportation system, if we take the passengers from Huaqiangbei in Futian District to Kexingyuan in Nanshan District as a transaction, then limiting the number of passengers from Huaqiangbei to Nanshan District is obviously not possible.

the number of passengers is obviously unlikely. Because in this process, passengers may pass through the subway, buses and other means of transportation, i.e., a transaction requires the cooperation of multiple modules to complete.

HPS

Hits Per Second, Hits Per Second. If each transaction completes one request, then TPS and HPS are equivalent. However, in a distributed scenario, multiple requests may be required to complete a transaction, so TPS and HPS are not equivalent.

Like the example in TPS, if a passenger needs only one direct bus ride from Huaqiangbei to Kexingyuan, then only one transportation participation (one request) is needed to complete the transaction.

QPS

Query Per Second, the number of client queries that can be responded to per second, is also an important measure of overall server performance.

Inside the transportation system, QPS can be understood as the sum of the number of passengers accommodated. Currently, most mainstream traffic limiting methods use HPS as a traffic limiting index.

2) Flow limiting methods

1. Traffic counter

The simplest and most direct way to limit the flow, for example, limiting the maximum number of requests to 100 within 5 seconds, ** access will be denied if the number exceeds this number**. Traffic counting in a transportation system limits the number of vehicles allowed to pass within a certain period of time.

There are two obvious problems with this method of limiting traffic.

centralized access

It is difficult to control the unit time, and it is easy to have centralized access. For example, the following scenario occurs:

!

In the first 4 seconds, there is only one access, in the 5th second, there are 99 accesses; in the 6th second, there are 99 accesses, and in the next 4 seconds, there is only 1 access. Globally, the 200 requests in 10 seconds is not exceeded, but the graph shows that this traffic situation is definitely an anomaly.

Unnecessary Traffic Limiting

There is a period of time when the traffic is exceeded, but it is not always really necessary to limit the traffic. For example, the following scenario:

If the two middle blocks of accesses happen to be in a 5-second cycle, then the traffic is exceeding the limit. In this case, the next 10 requests may be dropped, which doesn’t make sense.

2. Sliding Time Window

Sliding time window is a popular algorithm for limiting traffic, the main idea is to consider time as a forward scrolling window, as shown in the following figure:

!

It is characterized by the fact that ** time is processed in slices, and the sliding window counts the number of requests in a total time period each time. At the next time period, the previous time slice is discarded and the number of requests from the following time slice is added ** It solves the problems that can occur with traffic counters.

Its disadvantage is that it does not have a fine enough control over the traffic to limit the traffic concentrated in a short period of time and once the limit is reached, the requests are all violently rejected directly.

Inside the transportation system this may result in the loss of a portion of the passing traffic, which is not very friendly to the user experience.

3. Leaky Bucket Algorithm

The idea of leaky bucket algorithm is shown below:

!

A leaky bucket is a fixed-size queue that caches the requests sent by the client and then sends them evenly to the server.

If the client request rate is too fast and the queue of the leaky bucket is full, it will either reject it directly or go downgrade the processing logic without impacting the server side.

The advantage of the leaky bucket algorithm is that it is simple to implement and can use message queues to shave peaks and fill valleys. But it also has several problems:

• Leaky Bucket size is not easy to control, too large will bring more pressure to the server, too small may lead to a large number of requests are discarded;
• The rate of requests given to the server by the leaky bucket is difficult to control;
• The use of cached requests will make the response time of the request longer.

4. Token Bucket Algorithm

Token bucket algorithm and go to the hospital registration is almost the same logic, see the doctor before you need to register, and the hospital every day to put the number is limited:

!

Token bucket algorithm in the token is the same, the client before sending a request, need to first get the token from the token bucket. If it fetches it, it sends the request; if it fails to fetch the token, it can only be rejected and needs to wait.

The token bucket algorithm solves the three problems of the leaky bucket algorithm (the rate is difficult to control, the size of the leaky bucket is difficult to control and the problem of the request cycle becomes longer), and the implementation is not complex, the use of semaphores can be realized. In the actual flow limiting scenarios used in the most, such as Google and Guava in the use of token buckets to limit the flow.

Inside the transportation system, the license plate shaking is the token bucket flow limiting idea. Therefore, ** software systems serve life, software design also comes from life **.

5. How to limit flow in distributed scenarios

In distributed scenarios, can the above flow limiting scheme still be applied? Take a chestnut:

!

If we put the token into a separate middleware (e.g., Redis) for the whole distributed system, then the client needs to interact with the token bucket when invoking the combined service, and the combined service invokes the order, inventory, and account services, and the number of interactions obviously increases a lot.

One improvement is that the client first obtains four tokens before invoking the service, ** subtracts one token when invoking the combination service and passes three tokens to the combination service, and consumes one token when invoking the sub-services ** respectively.

In the transportation system, the distributed token bucket design is equivalent to a single transportation that needs to cross multiple hurdles, each of which requires a traffic limit token as a way to maintain the stability of the whole system.

4. Summary

For flow limiting, it is important to choose a suitable flow limiting algorithm. Token bucket algorithms have more obvious advantages, and they are also the flow limiting algorithms used in many large-scale projects.

When designing the system, these patterns need to be matched with the evaluation of business volume and performance test data for threshold modification, and these thresholds are best saved in a hot-updatable configuration center for real-time modification.

Interviewer: hmmm, good, besides flow limiting, are there any other methods that can be used in system design to maintain high availability?

** Flow limiting, fusing and service degradation are all important design approaches for high system availability**.

Fusing, the equivalent of putting a fuse between requests and services. When the service can’t carry the constant access pressure, the fuse breaks, preventing the server from crashing if it can’t carry the pressure.

Compared to flow limiting and fusing, service degradation is considered from a global system perspective. After a service fuses, it will generally let the request go through the processing method that implements the configuration, and this processing method is a degradation logic.

Interviewer: ok, interview time is almost up, let’s not expand on this for now, let’s do an algorithmic question ……

Due to the limited space, in this issue we only introduced the indicators and methods related to flow limiting.

1. introduction
2. map substructure
3. GET and PUT operations
4. DELETE operation
5. map expansion conditions
6. Considerations for using map
7. Postscript

1. Introduction

As a gopher (go as the development language of the program ape), every time in the face of the technical face of the Internet factory, will be asked to some go data structure or its unique mechanism of the underlying implementation.

For example, when I interviewed for the “Tencent TEG” department, an elegant and easy-going interviewer sitting across the video conference asked, “You usually use go language, right, so tell me what reference types are in go language?”

I thought, this is not difficult to me, so slowly answered: “ **go language reference types and slice (slice), pipeline (channel), function (func), interface (interface ) and dictionary (map) ** “.

The interviewer added, “I see that you have not worked for a long time, so I won’t ask you about the underlying mechanism of channel, which may be a bit difficult! In this way, you say a map of the underlying implementation, and map add and delete data and expansion of what operations”.

“Hey, I’m so grumpy, it means I can’t answer the question of channel, who am I looking down on?” I can’t help but think that this interviewer is so …… The bottom of the channel I understand the CSP, elegantly close the channel and other logic, if you ask about the source code implementation, I really will not!

However, the implementation mechanism of map in go is not very complicated, and there is a difference in design between other languages, and I spent time reviewing it before the interview. It seems that _the attentive interviewer is not only trying to find the most suitable candidate, but also the one who knows the most about them. _

2. map’s underlying data structures

So, I’ll describe the underlying implementation of a map in terms of its data structure, additions, deletions, and expansions.

First, we can find the underlying data structure code for map in the runtime/map.go package that comes with the go language:

1
2
3
4
5
6
7
8
9
10
11
12

type hmap struct {
     count     int    
     B         uint8  
     overflow  uint16

     buckets    unsafe.Pointer
     oldbuckets unsafe.Pointer  

     extra *mapextra
     ...

 }

Go maps are implemented as hmap structures. hmap records several attributes of a map, including the number of elements, the number of buckets [2^B], the addresses of the map’s buckets, and the addresses of the overflow buckets [we’ll talk about the concept of overflow buckets later, see below].

Among them, the structure of the extra overflow bucket is as follows, and the parameters record the address of the overflow bucket and the pointers to the upper and lower buckets, respectively:

1
2
3
4
5

type mapextra struct {
    overflow    *[]*bmap
    oldoverflow *[]*bmap
    nextOverflow *bmap
}

buckets records the address of the bmap that holds the map bucket (a bucket can be thought of as a bmap bucket). bmap is the key structure that records the actual data of the map, and its structure when it is not compiled is as follows:

1
2
3

type bmap struct {
    tophash [bucketCnt]uint8
}

tophash records the first 8 bits of the key in the map data, used to quickly compare whether the key exists in the map or not.

While the program is being compiled, the structure of bmap is as follows:

1
2
3
4
5
6

type bmap struct {
    tophash [8]uint8
    data    byte[1]  
    overflow uintptr
}

The overflow in bmap records the address of the next bmap, where overflow is of type uintptr instead of *bmap, in order to guarantee that the bmap does not contain any pointers at all, so as to minimize gc scanning.

But then the overflow bucket will be missing references, and may be gc’d when used because the element is empty, so hmap adds extra to store the pointer to the overflow bucket.

The data parameter stores the real map data (k-v key-value pairs), each bmap stores 8 k-v key-value pairs, and key-value pairs in the bmap are stored separately and consecutively as shown below:

When there are more than 8 k-v pairs, a new bmap is generated to store the data, and the overflow records the address of the new bmap.

summarize

In the go implementation of map, the structure represented is hmap, and hmap maintains a number of bucket buckets (i.e., bmap buckets, which are subsequently called buckets for the sake of uniformity). For intuitive understanding, we refine the key fields in hmap to graphical patterns:

Each element of the buckets array is a bmap structure. Each bmap bucket holds 8 k-v pairs, and if there are more than 8, a new bmap will be generated, and the overflow will record the address of the new bmap.

3. GET and PUT operations

After understanding the data structure of a map, we will learn how to access the data in a map.

3.1 GET Getting Data

Assuming that B=4, i.e. the number of buckets is 2^B = 16, suppose we want to get the value corresponding to key5 from the map.

1
2
3
4

m := make(map[string]string, 0)
...

fmt.Println(m[key5])

As shown above, the get process is divided into. 1:

1. Calculate the hash value of k5 [64-bit operating system, the result has 64 bits]. 2;
2. determine which bucket is in the last B bit of the **hash value **, 0100 is converted to decimal 4, so it is in bucket 4. 3. determine which bucket is in the first 8 bits of the hash value and the `tophash value;
3. according to the hash value of the first 8 bits and tophash for comparison, quickly determine in which bucket position, the next step. 4. the query key5 and the tophash to the next step;
4. compare the key5 of the query with the key5 of the bmap, and get the corresponding value5 if it matches exactly. 5. if no key5 is found in the current bucket, then get the corresponding value5;
5. if not found in the current bucket, go to the next overflow bucket connected by overflow, and repeat steps 3-4.

Repeat steps 3-4. 3.2 Storing Data in the PUT

The map assignment can be divided into the following steps (assuming we add key6):

• Determine which bucket it is by the last B bits of the hash value of key6;
• Iterate over the current bucket, compare the first 8 bits of the hash value of the key with tophash to prevent the key from being repeated, and then find the location where the elements are stored, i.e., insert the data at the first empty position of the bmap;
  ** If the current bucket is full, a new bmap bucket is created to insert the elements, then the address of the new bmap is recorded with overflow and a reference to the new bucket is added to the extra pointer array in the hmap.

3.3 About hash conflicts

**A hash conflict occurs when two different keys fall into the same bucket. When the bucket is not full, it is inserted using open addressing to find the first empty space from front to back. When the bmap bucket already has 8 k-v key-value pairs, a new overflow bucket (bmap) is created and the address of the new bmap is recorded by overflow and a reference to the new bucket is added to the hmap’s extra pointer array.

4. The DELETE operation

When a map is deleted:

• If the deleted element is a value type, such as int, float, bool, string, and array, the map’s memory is not automatically freed;
• If the element is a reference type, such as pointer, slice, map, and chan, part of the map memory will be freed, but the memory freed is the memory occupied by the reference type of the child element, and the memory occupied by the map itself is not affected;
• The memory will be released only after map is set to nil, and the memory will be reclaimed in the next GC.

So, ** map will not release memory immediately after deletion**, let’s verify this.

4.1 When map is a basic type

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46

var intMap map[int]int
var cnt = 8192
​

func printMemStats() {
    var m runtime.MemStats
    runtime.ReadMemStats(&m)
​
    log.Printf("Alloc = %vKB, TotalAlloc = %vKB, Sys = %vKB, NumGC = %v\n",
              m.Alloc/1024, m.TotalAlloc/1024, m.Sys/1024, m.NumGC)
}
​
func initMap() {
    intMap = make(map[int]int, cnt)
    for i := 0; i < cnt; i++ {
       intMap[i] = i
   }
}
​

func delMapKey() {
    for i := 0; i < cnt; i++ {
       delete(intMap, i)
   }
}
​
func main() {
    printMemStats()
​
    initMap()
    log.Println("after initMap, len(map) =", len(intMap))

    runtime.GC()
    printMemStats()
​
    delMapKey()
    log.Println("after delMapKey, len(map) =", len(intMap))
​
    runtime.GC()
    printMemStats()
​
    intMap = nil
    runtime.GC()
    printMemStats()
}

The final printout is:

Alloc = 108, TotalAlloc = 108, Sys = 6292, NumGC = 0
after initMap, len(map) = 8192
Alloc = 410, TotalAlloc = 424, Sys = 6867, NumGC = 1
after delMapKey, len(map) = 0
Alloc = 410, TotalAlloc = 425, Sys = 6931, NumGC = 2
Alloc = 99, TotalAlloc = 427, Sys = 6931, NumGC = 3

Where Alloc is the memory space occupied by the current heap object in KB, TotalAlloc is the cumulative space allocated for heap objects, Sys is the memory space occupied by the operating system, and NumGC is the number of garbage collection attempts. It can be clearly seen that ** when map deletes a basic type element, no matter how much GC is done, the memory is not freed. **

4.2 When map values are reference types

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

type Person struct {
    Name string
    Age  int
}
var intMap map[int]*Person
​
func initMap() {
    intMap = make(map[int]*Person, cnt)
    for i := 0; i < cnt; i++ {
       intMap[i] = &Person{
          Name: "zhangsan",
          Age:  20,
      }
   }
}

Change the value type of map to a reference type, leave the rest of the code unchanged, and run the result again:

Alloc = 108, TotalAlloc = 108, Sys = 6036, NumGC = 0
after initMap, len(map) = 8192
Alloc = 601, TotalAlloc = 615, Sys = 6867, NumGC = 1
after delMapKey, len(map) = 0
Alloc = 409, TotalAlloc = 615, Sys = 6931, NumGC = 2
Alloc = 99, TotalAlloc = 618, Sys = 6931, NumGC = 3

Comparing the two examples, we can find that when the map value is a basic type, the deletion of the key will not release space; while when the map value is a reference type, it will release part of the space (space on the heap of the value), but the memory occupied by the map itself is not reduced, why is this?

This is because the bottom of map will only increase or decrease the number of bmap buckets after ** applying for them**. The delete operation of the map will only set the data on the bmap to nil (if the value is of pointer type, the pointer object will be reclaimed in the next GC), but the memory occupied by the bucket itself remains unchanged, so the memory occupied by the map itself will not be reduced because of the deleted key.

In particular, if there is an empty space in the bmap bucket after a key is deleted, and a new key is added later, the empty space in the bucket may be filled**, and the memory footprint of the map remains unchanged.

4.3 How to Solve Memory Leakage Caused by a Map

When a map is frequently added or deleted, or too many keys are added (triggering a map expansion), even if ** these keys are deleted, the memory space will still be unavailable, thus triggering a memory leak. ** This can be solved by the following method.

We can use the following methods to solve the problem:

** Set the map that is no longer in use to nil, or restart the system periodically to allow the map to be reallocated;

• When the value of a map consumes too much memory, change the value to a pointer;
• Periodically copy the elements of the map to another map.

Generally speaking, most of the Internet businesses with high traffic impact are ToC scenarios, and the on-line frequency is very high. Some businesses may go online several times a day and restart and recover before the problem is exposed, which is not a big problem 🐶.

5. map expansion conditions

5.1 Expansion of the same capacity

Since map constantly puts and deletes elements, there may be a lot of intermittent empty bits in the bucket, which can cause bmap to overflow the bucket by a lot, resulting in longer scan times. So, this kind of expansion is actually a kind of collation, organizing the data of the backward bits to the front. ** With equal expansion, the elements are rearranged, but the buckets are not swapped, and the number of element buckets does not increase. **

5.2 Two-volume expansion

When the bucket array is not enough, the expansion will occur, then, ** elements will be rearranged, bmap bucket increase, the key elements inside the bucket may also be migrated **.

As shown in the figure, before expansion B = 2, after expansion B = 3, assuming that an element key hash value of the last three for 101, then by the introduction of the above can be seen, before expansion, hash value of the last two to determine the number of buckets. That is, ** before expansion, the last two digits of an element are 01, and the element is in bucket 1; after expansion, the last three digits of an element are 101, and the element needs to be migrated to bucket 5**.

5.3 Conditions for expansion to occur

1) Expansion condition 1: loading factor > 6.5

Under normal circumstances, if there are no overflow buckets, there are at most 8 elements in a bucket. When the average number of elements in each bucket exceeds 6.5, it means that the capacity of the current bucket is almost full and needs to be expanded.

loadFactor = number of elements in map / number of current buckets in map, i.e. loadFactor = count / (2^B)
From the formula, loadFactor is the average number of elements in each bucket in the current buckets.

2) Expansion Condition 2: Excessive number of overflow buckets

• When B < 15, if the number of overflowed bmaps reaches 2^B;

• When B >= 15, if the number of overflow’s bmap exceeds 2^15, then start to expand the capacity.

The reason for too many overflow buckets is that a large number of keys in the map have the same hash value after the B-bit, which makes individual bmap buckets keep inserting new data, which leads to longer and longer chains of overflow buckets.

Thus, when the map is being added, deleted, modified and checked, the scanning speed will become slower and slower. When expanding the capacity, you can rearrange the elements of these overflow buckets to make the position of the elements in the bucket more average, so as to improve the efficiency of scanning.

3) Details when expanding

1. In our hmap structure, there is an oldbuckets, when the expansion occurs, the old data will be stored in this first and then expand the buckets, at this time, the capacity of the buckets is twice the oldbuckets. 2. map expansion is an incremental increase in the capacity of the oldbuckets, so that the oldbuckets can be expanded;
2. map expansion is an incremental expansion, not a full expansion, and every time the map is deleted or modified, the operation of migrating from oldbuckets to buckets will be triggered. This is because full expansion with a large number of keys will consume a lot of resources and cause the program to stall. 3;
3. Before the migration is complete, each get or put traverses the oldbuckets before traversing the buckets.

6. Notes on map

1) Do not address elements

** As the elements of a map grow, the map bottom may reallocate space, resulting in invalid addresses before ** Let’s look at an example:

1
2
3
4
5
6
7
8
9
10
11
12
13

type Student struct {
    Name string
    Age  int
}
​
func f1() {
    m := map[int]Student{
        1: Student{Age: 15, Name: "jack"},
        2: Student{Age: 16, Name: "danny"},
        3: Student{Age: 17, Name: "andy"},
    }
    m[1].Name = "JACK"
}

A compilation error occurs in this case because map cannot be addressed. That is, you can get m[1], but you cannot make any changes to its value. If you want to modify value, you can use value with a pointer as follows:

1
2
3
4
5
6
7
8

func f2() {
    m := map[int]*Student{
        1: &Student{Age: 15, Name: "jack"},
        2: &Student{Age: 16, Name: "danny"},
        3: &Student{Age: 17, Name: "andy"},
    }
    m[1].Name = "JACK"
}

2) Thread unsafe

Suppose the number of barrels of a map is 4, i.e., B=2, and the number of current elements is also 4, i.e., the barrel is full. At this time, two goroutines (g1 and g2) read and write to this map, g1 inserts key1 and g2 reads key2, the following may happen:

1. g2 calculates the hash value of key2 [1101….. .101], B=2, and determines the bucket number to be 1;

2. g1 adds key1, which triggers the expansion condition, increasing B to 3 and expanding the number of buckets to 8. 3;

3. The key in the map starts to migrate, assuming that the migration is completed soon, and key2 is migrated from bucket 1 to bucket 5. 4. g2 is migrated from bucket 1 to bucket 5;

4. g2 traverses from bucket 1 and fails to get data!

So, when manipulating a map, you can use Go’s own sync.RWMutex locks, or use sync.Map (which supports concurrent and shared locks) to ensure thread safety.

7. Postscript

After answering these questions, the interviewer’s face remained unperturbed, and he said lightly, “Why don’t we move on to talk about channel?”

So, I disliked all the points I had learned, and the interviewer seemed to recognize my level, so he didn’t give me a hard time anymore and started the next topic.

After the interview, I couldn’t help but think: “It’s worthy of being a Penguin factory, it seems that the candidates are all strong!”

Nowadays, in the background of the Internet market is so cold, we want to meet the technical side of the big Internet factories, we must not only have a wide application of the technology we have learned, but also have to spend some effort on the cognition of the underlying mechanism.

After all, anyone can memorize the eight-stud text, anyone can brush the algorithmic questions, the interviewer in the screening also had to increase the difficulty, in order to test the candidate’s breadth and depth of knowledge! The gap between you and others may be the distance of a map.

1. 初始RPC
2. RPC与HTTP
3. 流行的RPC框架
4. Protobuf与gRPC
5. gRPC响应ChatGPT问答
6. 小结

1. 初始RPC

RPC 是什么？

RPC (Remote Procedure Call) is a computer communication protocol. The protocol allows a program running on one computer to call a subroutine in another address space (usually one computer on an open network), and the programmer calls it as if it were a local program, without having to additionally program for this interaction (no need to pay attention to the details). –Wikipedia

In layman’s terms, suppose there are two servers A and B, and two programs (program 1 and program 2) are deployed on these two servers. Since they are two machines, their IP addresses, memory space, etc. are definitely not shared, so how does program 1 call the methods of program 2?

At this point we need to agree on a protocol to allow applications on the two machines to communicate, RPC is such a protocol, it is through the following steps to allow the two programs to recognize the identity of the other program:

1. two machines to send and receive data, so one acts as a server, one acts as a client, they need to ** establish a TCP connection ** (on-demand call, can be a short connection, can also be a long connection);
2. before connecting a TCP connection, the client needs to know the IP address and port number of the server, where the IP address is a unique identifier of the host in the network and the port number is a unique identifier of the application program (aka process) on the host;
3. Before communicating, the server runs the application and listens for the corresponding process port number;
4. The client initiates an RPC remote procedure call, passing the parameters of the program interaction to the server, which then transmits them back to the client after processing the received data, disconnects the TCP connection and ends the call.

The whole process is shown in the following figure:

Translated with www.DeepL.com/Translator (free version)

• Client Stub (Client Stub): stores the server address message of the communication and packages the client’s request ** into a message body that can be transmitted in the network;

• Server Stub (server-side stub): receives messages sent by the client and packages the return results into message bodies that can be transmitted across the network.

• Sockets (network sockets): a set of program interfaces for applications that can be used to exchange data between different hosts in a network.

Translated with www.DeepL.com/Translator (free version)

2. RPC vs. HTTP

**HTTP & RPC, how to choose? **

After learning about RPC, some people may still wonder: since they are both communication protocols, should we choose HTTP (HyperText Transfer Protocol) or RPC protocol for program interaction and application development?

This starts with the attributes of both of them. First, the transfer protocol:

• RPC is a communication protocol based on the TCP transport layer or the HTTP2 application layer;
• HTTP is based on HTTP protocols only, including HTTP1.x (i.e. HTTP1.0, 1.1) and HTTP2, and many browsers currently use 1.x by default to access server data.

Performance consumption (in terms of data type comparison):

• RPC, can be based on gRPC (an RPC framework) to achieve efficient binary transmission;
• HTTP, most of which is implemented through json, byte size and serialization are more performance consuming than gRPC.

On load balancing:

• RPC, basically comes with a load balancing strategy;
• HTTP, need to configure Nginx, HAProxy to realize.

Transfer efficiency:

• RPC, the use of custom TCP protocol, you can make the request message size smaller, or use the HTTP2 protocol, can also be very good to reduce the size of the message, improve the transmission efficiency;
• HTTP, if it is based on HTTP1.x protocol, the request will contain a lot of useless content; if it is based on HTTP2.0, then a simple encapsulation can be used as RPC, which is the standard RPC framework is more advantageous for service governance.

In summary, we can easily find that **RPC from the performance consumption and transmission efficiency, as well as load balancing and other aspects are stronger than HTTP **. At this point, careful friends may have found that, that why our common systems and websites are using the HTTP protocol, do not change to RPC communication?

To give a common example, HTTP is like Mandarin, RPC is like a local dialect, such as Cantonese, southwest of Yunnan, Guizhou, Sichuan.

The advantage of speaking Mandarin is that everyone understands it, and most people speak it, so HTTP has a certain universality. Speaking dialect, the advantage is that it can be more concise, more confidential, more customizable, the disadvantage is that the other party who “speaks” the dialect (especially the client side) must also understand, and once everyone speaks a dialect, it is difficult to change the dialect. So RPC is generally used for internal service calls, such as between service A and service B in the Ali Taobao system.

3. Popular RPC frameworks

There are many popular RPC frameworks, here are three common ones.

1. gRPC: gRPC is an open source project announced by Google in 2015, based on the HTTP2.0 protocol, and supports many common programming languages. The HTTP 2.0 protocol is an upgraded version of the binary-based HTTP protocol, which supports features such as multiple concurrent data transfers.
2. Thrift: Thrift is an internal system cross-language RPC framework developed by Facebook, which was contributed to the Apache Foundation in 2007 and has become one of the many open source projects of Apache.
3. Dubbo: Dubbo is Alibaba in 2011, an open source RPC framework , in many Internet companies and enterprise applications are widely used , provides a series of protocols and serialization framework , pluggable , but only supports the Java language.

Foreign RPC evaluation, based on the comparison of the test situation of each RPC framework, from the ** throughput rate, response time and stability **, gRPC comprehensive performance is better, but also a lot of domestic companies in the use of the RPC framework. Moreover, gRPC is realized by the go language, with the popularity of microservices, cloud computing, go language companies and projects are increasing, so gRPC has also become the go language internal system communication choice.

gRPC based on **ProtoBuf (Protocol Buffer) serialization protocol ** development, its principle is through the IDL (Interface Definition Language, Interface Description Language) file to define the parameters of the service interface and the type of return value, and then through the code generation tool to generate the template of the server and client code. In this way, we only need to implement an IDL file and business interaction code, and then we can use gRPC to communicate.

A diagram shows the difference between gRPC and HTTP:

4. Protobuf and gRPC

4.1 Introduction to Protobuf

Proto Buffer protocol (protobuf for short, same as below), like json and xml, is a kind of data serialization (serialization, that is, converting a piece of data in memory into binary form, and then transferring or storing it over the network).

• protobuf is a cross-language, cross-platform serialization protocol;
• Not only limited to gRPC, protobuf can also be used for data transfer and storage in other scenarios;

Unlike json and xml, protobuf needs to define IDL (Data Definition Rules) when using it, which has the advantage of smaller data size and faster transfer speed.

The transport protocol of gRPC uses protobuf, so we need to learn the rules of writing protobuf files first.

4.2 Protobuf Defining Data Structures

Similar to yaml and xml files, protobuf files need to be written in a specific format, and the following is a generalized way to write a protobuf file [gpt.proto]:

In addition to the description in the image above, protobuf files also have some key fields, such as message, which is the most basic type in the protobuf protocol, equivalent to a class object in Java or a struct in Go. As you can see in the figure above, each message has one or more fields and field types, which are equivalent to the parameters and parameter types of an object.

Once we’ve written the protobuf file, we can start writing the communication logic for gRPC.

4.3 gRPC implementation

As we said above, gRPC is a framework for cross-language communication, so the server and client can be different languages. Next, let’s demonstrate the process of implementing gRPC communication in go language.

Steps:

1. write protobuf file
2. generate Go code
3. write the client side to listen on the port
4. write the server side, request data

First, let’s create a new project, the directory structure is as follows [wecom project, is used to do GPT interaction, you can imitate, the important folders and file names have been circled in red]:

1) Writing the protobuf file

According to the above protobuf rules, we first write the protobuf file used in this project [protos/gpt/gpt.proto].

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

syntax = "proto3"
​
option go_package = "./;gpt"
​
package gpt
​
service Greeter {
  rpc GetGPTMessage (GPTRequest) returns (GPTReply) {}
}
​
message GPTRequest {
  string content = 1
}
​
message GPTReply {
  string message = 1
}

2) Generate Go code

We need to use protoc tool to generate Go language code, first download the toolkit [according to different computer systems to install proto package]:

• Windows 64位 点这里下载
• Mac Intel 64位 点这里下载
• Mac ARM 64位 点这里下载
• Linux 64位 点这里下载

Then install the golang plugin [generating plugins for other languages such as Java and Python is different, see the official gRPC documentation for details]. [https://doc.oschina.net/grpc?t=58008)】

go install google.golang.org/protobuf/cmd/protoc-gen-go@latest

This completes the installation of our toolkit.

After the installation is complete, go to the directory where the proto files are located:

cd protos/gpt/

Generate code in Go [if protoc does not exist, the protoc toolkit is not installed or there is a problem with the environment variable settings].

protoc –go_out=. –go_opt=paths=source_relative –go-grpc_out=. –go-grpc_opt=paths=source_relative gpt.proto

At this point, there are 3 files in the current directory [protos/gpt]:

gpt.pb.go
gpt.proto
gpt_grpc.pb.go

3) Add dependencies

If, like me, there is code marked in red after generating the code, you can add the dependency in the main directory [/wecom]:

go mod tidy

If there is still red after go mod tidy, it may be caused by a lower version of golang’s dependencies, so you need to change golang’s dependencies in go.mod:

go 1.18

4) Write business code to implement the server side

First, import the grpc plugin package in your project

go get google.golang.org/grpc

Then write the server-side business logic [gpt_server/main.go

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38

package main
​
import (
    "context"
    "flag"
    "fmt"
    "google.golang.org/grpc"
    "log"
    "net"
    pb "wecom/protos/gpt"
)
​
var (
    port = flag.Int("port", 50051, "port")
)
​
type server struct{
    pb.UnimplementedGreeterServer
}
​
func (s *server) GetGPTMessage(ctx context.Context, in *pb.GPTRequest) (*pb.GPTReply, error) {
    return &pb.GPTReply{Message: "gpt response"}, nil
}
​
func main() {
    flag.Parse()
    list, err := net.Listen("tcp", fmt.Sprintf(":%d", *port))
    if err != nil {
        log.Fatalf("listen failed, %v", err)
    }
    s := grpc.NewServer()

    pb.RegisterGreeterServer(s, &server{})
    log.Printf("listen success, %v", list.Addr())
    if err := s.Serve(list); err != nil {
        log.Fatalf("server failed, %v", err)
    }
}

Run the main function:

Next, we implement another client binding to a port to request messages from the server.

5) Client-side Logic

From the server-side implementation above, the gRPC implementation is very simple, just follow the template generated by protobuf to fill in the business code! This process, we only need to focus on the server-side and client-side connection communication, and their connection is not very deep, and our HTTP listening and binding is the same principle.

Client-side business code [gpt_client/main.go

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

package main
​
import (
   "context"
   pb "dm-lite/resource/proto/gpt"
   "flag"
   "google.golang.org/grpc"
   "google.golang.org/grpc/credentials/insecure"
   "log"
)
​
const defaultName = "world"
​
var (
   addr = flag.String("addr", "localhost:50051", "")
   name = flag.String("name", defaultName, "")
)
​
func main() {
   flag.Parse()
   conn, err := grpc.Dial(*addr, grpc.WithTransportCredentials(insecure.NewCredentials()))
   if err != nil {
      log.Fatalf("Dial failed, %v", err)
   }
   defer conn.Close()
   c := pb.NewGreeterClient(conn)
   ctx := context.Background()
   r, err := c.GetGPTMessage(ctx, &pb.GPTRequest{
      Content: "hello",
   })
   if err != nil {
      log.Fatalf("GetGPTMessage failed, %v", err)
   }
​
   log.Printf("get reply: %v", r.GetMessage())
}

When the server is listening on port 50051, we can run the client to call the GetGPTMessage method of gRPC. Run the client main function to get the result:

POTUS, the gRPC interface call was successful!

5. gRPC response to ChatGPT Q&A

Streaming RPC

The above implements the gRPC interface for real-time response, i.e.: a simple pattern of one question and one answer. If compared to a scenario in an interview, the simple pattern looks like this:

(Interviewer) Q: Do you know gRPC?
(Candidate) A: Yes, gRPC is an RPC framework initiated by Google;
(Interviewer) Q: What else?
(Candidate) A: gRPC is based on HTTP/2 protocol transport;
(Interviewer) Q: What else?
(Candidate) A: It uses Protocol Buffers as the interface description language;
(Interviewer) Q: Can we finish this at once?
(Candidate) A: ……

Then the interviewer asked us to do it, so we can’t afford not to do it, so the streaming pattern RPC appears:

(Interviewer) Q: Do you know gRPC?
(Candidate) A: Yes, gRPC is an RPC framework initiated by Google… It is based on HTTP/2 protocol transport… And it uses Protocol Buffers as the interface description language.
(Interviewer) Thinks to himself: not bad! I’m more of a stick-in-the-mud kind of guy than a toothpaste-squeezing kind of guy!

Next, we add a client-side streaming RPC interface to the proto file:

1

rpc GetGPTStreamData (GPTRequest) returns (stream GPTReply) {}

5.1 添加流式接口

Improvement of protobuf files [gpt/gpt.proto

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

syntax = "proto3"
​
option go_package = "./;gpt"
​
package gpt
​
service Greeter {
  rpc GetGPTMessage (GPTRequest) returns (GPTReply) {}
  rpc GetGPTStreamData (GPTRequest) returns (stream GPTReply) {}
}
​
message GPTRequest {
  string content = 1
}
​
message GPTReply {
  string message = 1
}

Following the three-step strategy for gRPC development, we’ll start by generating template Go code from a proto file:

protoc –go_out=. –go_opt=paths=source_relative –go-grpc_out=. –go-grpc_opt=paths=source_relative gpt.proto