Trace-Based Anomaly Detection with Contextual Sequential Invocations

Nowadays, microservice architecture has been widely adopted in various real systems because of the advantages such as high availability and scalability. However, microservice architecture also brings the complexity of operation and maintenance. Trace-based anomaly detection is a key step in the troubleshooting of microservice systems, which can help to understand the anomaly propagation chain and then locate the root cause. In this paper, we propose a trace-based anomaly detection approach called TICAD. Our core idea is to group the invocations according to their microservice pairs and then perform anomaly detection individually. For each distinct microservice pair, we propose a neural network based on LSTM and self-attention to automatically learn the contextual pattern in the target invocation and previous invocations. Detected invocation anomalies can be further used to infer the trace anomalies. We have verified it on a public data set and the experimental results show that our proposed method is effective compared to the existing approaches.