Shulou(Shulou.com)06/03 Report--

The annual Singles Day holiday is not only a carnival of minced hands, but also a stage for testing the technical level and technological innovation practice of major e-commerce technical teams, constantly setting new highs in sales, trading and payment peaks. these amazing figures are inseparable from the strong technical support. IT168 hopes to reveal to readers the huge investment and technological innovation of major e-commerce platforms behind the "super project" of Singles Day, in the form of technical reports, so that more people can understand and respect technology, promote technology exchange and sharing among the same industry, and promote the improvement of the overall technological level of the industry.

Summary of this article:

The Singles' Day promotion in 2017 is a "big parade" of JD.com 's fourth retail revolution. This "military parade" JD.com made a large-scale layout of the offline market, including more than 160 JD.com House and JD.com stores, nearly 200 JD.com mother and child experience stores, more than 1700 Jingdongbang service stores, and more than 5000 JD.com home appliance stores. Wal-Mart has more than 400 stores across the country, nearly 10,000 stores of JD.com co-brands and hundreds of thousands of convenience stores connected to JD.com 's handheld cabinet treasure.

After many trials and tests over the years, JD.com trading platform has obviously been good at dealing with ultra-high concurrency and super-large traffic. This year's Singles' Day, JD.com focused on refinement and artificial intelligence, upgrading the technical support, GMV, user experience and other aspects of intelligence.

Author: Wang Xiaofeng

About the author: Wang Xiaofeng, JD.com Mall trading platform architect, responsible for trading platform infrastructure planning, has rich experience in dealing with high traffic, high concurrency, low latency, high availability system design.

Text:

After many trials and tests over the years, JD.com trading platform has been adept at dealing with the technical guarantee of promoting system stability with ultra-high concurrency and large traffic. On the whole, the system stability mainly starts from two aspects: high performance and high availability.

JD.com Mall trading platform provides basic core services such as users, goods, inventory, prices, promotions and coupons, platform services for gold trading processes such as shopping carts, settlement pages, order centers, and omni-channel services such as PC, APP, Wechat, Mobile QQ, Kepler, etc.

In the past year, we have mainly ploughed and refined and upgraded intelligently in many aspects, such as technical guarantee, GMV upgrading, user experience and so on. In this paper, we share from two aspects of technical support and technology-driven business.

Technical guarantee of JD.com 's Singles Day Super Project

Dynamic load balancing and dynamic current limiting are two major applications of intelligent traffic allocation.

At present, intelligent traffic allocation is mainly used in two aspects, one is dynamic load balancing, the other is dynamic current restriction.

The load balancing algorithm is mainly based on random and polling. The specifications of the expanded servers may vary from year to year, and the hardware performance is high and low. When forming a large cluster, both random and polling will have a bucket effect. That is, the performance of the entire cluster is determined by the server with the lowest hardware specification.

We used to "save the nation" by pre-configuring weights or classifying machines with the same specifications. This method is feasible when the cluster size is small, but with the continuous increase of the cluster size, this process not only becomes time-consuming and laborious, but also very fragile, error-prone, and more difficult to deal with manually in the era of large-scale containerization. Therefore, we urgently need load balancing to dynamically identify the carrying capacity of the server and automatically adjust its weight.

The purpose of traffic restriction is to limit the flow rate to a reasonable range that the system can bear when there is a sudden increase in traffic, so that the system will not be destroyed by instantaneous flow. Through the full-link pressure test, we have been able to accurately estimate the system expansion requirements, but everything is in case, so we still need to be prepared to limit the current. Common current-limiting dimensions are traffic and concurrency, as well as smooth current-limiting algorithms, such as leaky bucket (Leaky Bucket) and token bucket (Token Bucket) algorithms, which are usually used together.

These policies and algorithms are usually statically configured according to a security threshold estimated in advance, but the actual running environment is often very complex and changeable, and the system may not be able to handle the visit volume and concurrency before reaching the security threshold. At this time, it is too late to manually intervene to adjust the threshold. Therefore, we also need the current limiter to dynamically identify the carrying capacity of the server and automatically adjust the threshold.

Through the comprehensive calculation of CPU utilization, CPU Load, number of TCP connections, response delay and other system and application performance indicators, we can calculate the server load capacity and health status in real time, feedback to the load balancer and current limiter in real time, and achieve intelligent allocation of traffic, which can not only maximize resource utilization, but also give it sufficient security protection.

From self-directed and self-conducted exercises to confrontational military exercises and then to self-service drills, JD.com 's fault drills are upgraded again.

When preparing for the Singles Day last year, our fault exercise was upgraded from "self-directed and self-performed" to "confrontation military exercise," which was divided into red and blue offensive and defensive forces, which were respectively responsible for fault recovery and fault manufacturing. The Blue Army independently designed the exercise subjects, which the Red Army did not know in advance and would only be told to launch an "attack" within a certain period of time. The headquarters requires that the point of failure must be identified within 5 minutes, the plan must be accurately implemented, and the system must be effectively restored in order to pass the examination.

This year, our fault exercise has been directly upgraded to self-service. The Blue Army can choose its own target applications and target clusters in the military exercise system. It can randomly select target machines, combine multiple applications and clusters, and combine a variety of faults, including network packet loss, port failure, soaring CPU, memory, disk utilization, Docker instance downtime, Redis instance downtime, and so on. It can also be carried out regularly, or even release a "smoke bomb" through a false alarm. This not only solves a large number of manual intervention and interaction, but also makes the fault cases more random and real, and can more truthfully test the feasibility and completeness of the plan. It greatly tests the psychological quality of the emergency team in dealing with faults and the response ability to deal with random events.

Technology-driven business, intelligent upgrade of JD.com trading platform

The success of Singles Day promotion is also inseparable from the support of big data and intelligent algorithms, and the strong product design and system research and development capabilities of the trading platform play an important role in this "big military parade". The trading platform continuously upgrades the data intelligently on the original business system; at the same time, on the basis of the original trading product line, it has launched intelligent marketing product lines for user experience, brands and sales.

Choose elves wisely to make your goods the best value for money.

Smart selection wizard belongs to the intelligent upgrade of the basic product line, and its first phase mainly includes two products, one is that the shopping cart selects the best promotion for user intelligence, and the other is that the settlement page selects the best coupons for user intelligence.

JD.com marketing activities, whether promotions or coupons, there will be a SKU or multiple SKU can meet multiple promotions or coupons at the same time, in addition to category dimension, merchant dimension, and vice versa. Therefore, no matter seeking the best coupon or the optimal promotion, it is actually a process of finding the optimal solution of the combination.

When the amount of data is small and the rule set is small, the combined result set is relatively small, the computational complexity is low, and even without system calculation, the user can identify it by himself, and there is no problem of user experience. When the amount of data is large, especially as JD.com currently has a huge amount of SKU, shopping carts can add hundreds of SKU at the same time, a variety of promotional rules, both full reduction and free gifts, promotions and coupons can also be superimposed, each user can receive thousands of coupons. At this point, the system is faced with a dilemma: if you cannot make the best choice for the user, the user will have to make a "brainstorm" when shopping; if you want to do it, it will be a massive computing process to search for the best discount directly and violently, and it will cost a lot of money. It may not be figured out yet, and the front-end call has already timed out.

At this time, the role of intelligent algorithm is highlighted.

Mathematically, there is always an optimal solution for any given random problem. The task of intelligent algorithm is to find this solution, which is equal to or close to the optimal solution in mathematics to the greatest extent. In other words, our work is a bit like target practice, the center of 10 rings is a complete hit, but 9.999 rings is also a good result. The 9.999 ring means that if the theoretical optimal value is 1000 yuan, we have calculated to 999.9 yuan. More importantly, for larger problems, the optimal solution is always unknown, and it may take days, years, or even tens of thousands of years to adopt an exhaustive approach.

Through the continuous tuning of intelligent algorithms, the accuracy of smart promotion and smart coupons ranges from 95% to 100%, and the most conservative probability is 97.2%. At the same time, the performance can be maintained within 5 milliseconds, greatly improving the user experience and greatly improving the conversion rate.

Smart marketing, firmly targeting potential users

In addition to the intelligent upgrading of the original functions, the trading platform has also launched intelligent systems that include JD.com exclusive, customer acquisition and repurchase artifacts, and so on, which are deeply combined with the trading platform, greatly improving the flexibility, output and efficiency of procurement and sales operations. In the past, purchasing and selling can only be combined into the user package that you want to market through business experience, selecting the corresponding user tags and user portraits, and then reaching the user through text messages or pages. This process relies heavily on the business judgment of purchasing and selling. Once there is a deviation in the judgment, it will cause a deviation in the effect.

It is based on this business demand that the trading platform has launched a smart marketing product line.

Take JD.com 's exclusive use of this product as an example. As soon as this product was launched, it aroused great repercussions and enthusiasm in procurement and sales. Among them, the most eye-catching is the inner core of this product, "high-potential user model", which is based on user, commodity and behavior data. The modeling team uses the technology of data mining and machine learning algorithm to build a prediction model for users to buy goods, and output high-potential users and target goods.

This logic seems simple, but it is very difficult to achieve high-precision accuracy. Because the behavior of users in the shopping process will be affected by a variety of factors, including prices, goods, channels, promotional activities and so on, the model needs high-frequency training in order to adapt to the variety of the market.

At present, the accuracy of the high potential user model for category identification is as high as 80%, of which the prediction accuracy of rigid demand categories such as smoke stoves and water heaters can reach more than 85%. For the purchase of SKU prediction accuracy of more than 50%, relatively simple categories of tobacco stoves and other categories, SKU prediction accuracy can reach 80%. Basically achieved "guess what the user wants to buy, the user will buy what". Over the past two years, the Gaoqian user model has continuously expanded the category of forecast, and has been more and more used in daily and promotional activities.

In addition to the accurate data model, the customer-grabbing system is also the most effective "smart promotion tool". By connecting different promotion methods, such as tokens, coupons, reservation pre-sale, second kill, etc., purchasing and selling can flexibly use high potential user data and customize the optimal promotion scheme for each user, that is, different users match different products, apply different prices, and achieve a real thousand people and thousands of people.

During the non-promotion period, purchasing and sales can use intelligent models and systems to more accurately locate high-potential people, and establish exclusive promotion and exclusive access in the operation of the system, so that users can feel JD.com 's differentiated channel advantages. to achieve the goal of seizing those users who "browse in JD.com but are about to place orders in other channels", and retain JD.com 's high-probability buyers who are about to lose.

During the promotion period, due to the fierce competition between channels, many users browse price comparison across channels, which is very easy to lose; at the same time, procurement and sales have higher requirements for the flexibility of promotion methods in order to respond to the marketing strategy of competitive channels in time. At this time, the intelligent promotion system around users is very necessary. Through the intelligent system, purchasing and selling can release the locked promotional goods and promotional prices (such as reserved goods, etc.) to the identified high-potential users in advance, helping brands to lock users firmly. At the same time, the smart system also gives more opportunities for flexible operation in order to meet the more diversified market needs.

Artificial Intelligence has become the New main Force of JD.com 's Technology

Artificial intelligence is surging in an unprecedented manner, rapidly entering people's field of vision. JD.com has been committed to using technology to drive business growth and improve user experience in an all-round way. With the comprehensive application of artificial intelligence and other technologies in JD.com, more and more problems solved by sea of people tactics in the field of business and technology will be gradually taken over by machines, allowing technicians to exert greater creativity, while ensuring the stability of the system. Improve operational efficiency, significantly reduce costs, and enhance user experience.

Preview of the next phase of the "Singles' Day" super project series:

In the data era, big data computing has penetrated into a variety of industries, business precipitation data, data computing produces new business value, big data computing is constantly using this way to promote business development. So in the face of such a business scene with high concurrency, high traffic and unique characteristics, what opportunities does big data have to show his talents in real-time computing? Please look forward to the next issue of NetEase Yun senior technical experts to bring the "NetEase double 11" Super Project: how can big data's real-time computing be tailored for you? "share.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.