Distributed tracing
I work on using distributed traces to understand latency and errors in microservice systems. The work combines analysis of individual requests with patterns across many traces.
CRISP
With colleagues at Uber and UC Santa Barbara, I worked on CRISP, a tool for critical path analysis of microservice request traces.
A request can involve many services running in parallel. CRISP identifies the chain of dependent operations that determines when the request finishes. We use this analysis to find latency bottlenecks, identify operations that affect many requests, and investigate performance problems.
- Paper — USENIX ATC 2022
- Talk recording
- Code
- CRISP: Critical Path Analysis for Microservice Architectures — Uber Engineering
- Data
Errors in microservices
I also worked with colleagues on The Tale of Errors in Microservices. We studied RPC errors that don’t cause the overall request to fail, and how they still add latency. We used distributed traces to identify these errors and estimate how much latency could be reduced by fixing them.