User case of OpenMLDB on anomaly detection of transaction system in AIOps field

Today we are going to introduce OpenMLDB’s user cases in the field of AIOps. AIOps is short for artificial intelligence for IT operations.

(https://github.com/4paradigm/OpenMLDB)

Transaction system is the core system of most enterprises. In order to ensure its continuous and stable operation, its health has become the core indicator of operation and maintenance health. Today, we mainly share the two indicators of transaction success rate and TPS to automatically detect whether the transaction systemis abnormal. The specific definition of the indicators is:

1. Transaction success rate, which counts the proportion of transaction success every 10 seconds as a cycle
2. TPS, here we count the average TPS every 10 seconds

Transaction system anomaly detection service architecture

The whole detection step is divided into:

Step1. TPS transaction volume and success rate statistics are generated in the transaction system cycle (every 10s);

Step2.Write the statistical data into the database, and then make an algorithm model reasoning, and store the reasoning results;

Step3. After completing the machine learning reasoning, conduct a rule to judge whether the decision is to alarm

How does the transaction system anomaly detection system deal with system jitter?

Because the transaction system may have jitters due to various problems, which of these system jitters needs to be paid special attention by the operation and maintenance personnel, which of these jitters can be swallowed by the system, these judgement need to be flexible to solve these problems. The whole anomaly detection system solves these problems from the following two aspects:

1. A variety of algorithms are combined for reasoning. Each algorithm performs reasoning, and the results are weighted. The greater the weight, the higher the feasibility;
2. Human makes decision rules and judges whether to alarm according to peoples’ experience. For example, check the reasoning results of the last 6 times, if the weight of 5 of the 6 times is greater than a certain value, the alarm will be triggered.

System architecture of transaction anomaly detection based on OpenMLDB

With high real-time performance, anomaly detection is required to quickly determine whether there is an anomaly to prevent the transaction.

Algorithm model freedom allows data scientists to give full play to data value and improve decision-making efficiency. This is also the core advantage of OpenMLDB. It reduces the technical threshold to develop a machine learning application. The system capacity can support 10 times the capacity growth of monitoring objects every year.

The github link of OpenMLDB: https://github.com/4paradigm/OpenMLDB