The Key Techniques Of Root Causes Location In Multi-dimensional Time Series Data

The quality of services(QoS)guarantee is extremely important in Internet companies,especially the rapid location of root causes and the fault recovery when sudden service anomalies appearing,which is one of the basic capabilities that Internet companies need to have.To this end,companies need to perform real-time monitoring of services and systems,and regularly collect monitoring indicators of various dimensions.When the service is abnormal,the operation and maintenance personnel can analyze the collected data to locate the root cause.However,the service scale of the current Internet giants is huge and the business scenarios are diverse.There are many indicators that need to be tracked and monitored,and the amount of monitoring data collected is huge.How to quickly locate the root cause combination from the huge search space generated by massive multi-dimensional monitoring indicators has become a very challenging problem.To sovle this problem,some automatic root cause location methods have been proposed to replace the tedious manual investigation.The main idea is to build a root causes tree based on monitoring indicators.From coarse-grained single-dimensional indicators to fine-grained multi-dimensional indicator combinations,anomaly detection and root causes judgment are performed in sequence.This process can also be viewed as a traversal of the root cause tree from top to bottom.Although this type of method can effectively find the combination of root causes,it also has certain defects,for example: only the root cause of large-scale anomalies can be located,and the small-scale but important anomalies are ignored;Data;need to adjust parameters frequently,or the timeliness is too poor,etc.In order to make up for the shortcomings of the existing methods,this paper proposes a new idea of root cause location.Unlike existing methods,we traverse the root cause tree from the bottom up to locate the root cause.The main idea is to reduce the root cause search space through anomaly screening of multi-dimensional index data;then combined with the statistical division method based on the root cause tree to quickly and accurately locate the root cause combination.The main work of this article includes the following two aspects:(1)We propose a bottom-up statistical root cause location algorithm named Bitdict.In order to overcome the shortcomings of existing algorithms,the Bitdict algorithm uses the bottom-up search strategy based on the root cause tree principle to search for root cause combinations.First,we detect whether the most fine-grained leaf nodes in the root cause treeare abnormal.Then,the statistical division is performed based on the frequency of occurrence of each attribute in the abnormal leaf node.Finally,the root cause attribute combinations are located one by one hierarchically.(2)In order to make better use of the Bitdict algorithm,a prototype of root cause location system is designed.In the system,we combine the prediction method based on moving average and the automatic threshold selection method to process the abnormal classification of leaf nodes.A more accurate set of abnormal leaves is provided as the input of the Bitdict algorithm,thereby improving the efficiency and accuracy of root cause location.In the experiments,we use public data sets.The experimental results show that the proposed algorithm Bitdict has excellent performance of root cause localization and strong robustness.Moreover,we can obtain better performance by designing a specific root cause localization system.