Performance Diagnosis in Cloud Microservices Using Deep Learning

Abstract

Microservice architectures are increasingly adopted to design large-scale applications. However, the highly distributed nature and complex dependencies of microservices complicate automatic performance diagnosis and make it challenging to guarantee service level agreements (SLAs). In particular, identifying the culprits of a microservice performance issue is extremely difficult as the set of potential root causes is large and issues can manifest themselves in complex ways. This paper presents an application-agnostic system to locate the culprits for microservice performance degradation with fine granularity, including not only the anomalous service from which the performance issue originates but also the culprit metrics that correlate to the service abnormality. Our method first finds potential culprit services by constructing a service dependency graph and next applies an autoencoder to identify abnormal service metrics based on a ranked list of reconstruction errors. Our experimental evaluation based on injection of performance anomalies to a microservice benchmark deployed in the cloud shows that our system achieves a good diagnosis result, with 92% precision in locating culprit service and 85.5% precision in locating culprit metrics.

Artificial Intelligence for IT Operations (AIOPS) Workshop White Paper

“…Performance Diagnosis in Cloud Microservices using Deep Learning Wu et al [22] introduce an approach to locate the culprits of microservice performance degradation, by investigating the cloud infrastructure on application level and machine level. They mine a service dependency graph which includes the machine and service level of the cloud infrastructure.…”

Enhancing the Analysis of Software Failures in Cloud Computing Systems with Deep Learning

“…[24] introduced an anomaly detection approach that leverages an online clustering method to define the normal behavior of monitored components. Wu et al [25] proposed a method that applies a dependency graph and an autoencoder to identify the causes of the performance degradation in the microservices of the cloud. Both previous works evaluated the proposed solutions by injecting performance anomalies in the cloud computing system.…”

Anomaly Detection and Failure Root Cause Analysis in (Micro)Service-Based Cloud Applications: A Survey

“…The baseline model is then used online to detect whether the newly monitored KPIs on a service diverge from the baseline model, hence indicating that the service suffers from some anomaly. This technique is adopted by Gulenko et al [27], MicroRCA [91], Wu et al [90], LOUD [51], and DLA [75], which train the baseline model by considering the KPIs monitored in normal runs of the application, viz., assuming that no anomaly occurred in such runs.…”

“…Differently from the above discussed, "KPI-agnostic" techniques [27,51], MicroRCA [91], Wu et al [90], and DLA [75] focus on given KPIs. MicroRCA [91] and Wu et al [90] consider the response time and resource consumption of application services, which they monitor by requiring the Kubernetes (k8s) deployment of multi-service applications to be instrumented as service meshes featuring Istio [31] and Prometheus [68]. The latter are used to collect KPIs from the application services.…”

“…Various solutions have been proposed to automate the detection of failures and to automatically determine their possible root causes. Existing solutions for failure detection rely on identifying anomalies in the behaviour of services, which can be symptoms of their possible failures [58,60,90]. Once an anomaly has been detected in a multi-service application, further analyses are enacted to determine the possible root causes for such an anomaly [42,77].…”