CRISP: Critical Path Analysis of Large-Scale Microservice Architectures
summary
CRISP performs critical path analysis over Jaeger traces across ~40K endpoints at Uber, inferring last-return sinks despite clock skew and missing spans. It constructs per-trace critical paths, produces top-down differential flame graphs and bottom-up impact rankings, and feeds compact feature vectors to an autoencoder anomaly detector. Results: finds five real optimizations, reduces anomaly false positives up to 50%, speeds training up to 28× and inference up to 67×, and shrinks per-trace endpoint sets about 10×. CRISP critical path
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:29What is microservice architecture?
- 1:27Why microservices?
- 1:59Microservice Challenges: Complexity
- 2:57Distributed tracing
- 5:01Our solution
- 5:59Critical Path Analysis (CPA)
- 6:57CPA example
- 8:13Challenges applying CPA on real-world traces
- 9:03Critical Path on Perfect Traces
- 9:33Critical Path on Real Traces
- 11:01Design of CRISP (Critical path and Span)
- 11:49Top-Down Analysis
- 12:16Differential Analysis
- 13:02Bottom-Up Analysis
- 14:42Anomaly detection
- 16:17Recall Improvement
- 17:20Training and Inferencing Speedup
- 17:59Conclusion
related talks
04mCritical Path-driven Development
32mA Picture is Worth 1,000 Traces
21mNavigating the Complexities of Service to Service Invocations: Deep and Brief Dive Into Causality
51mNavigating Distributed Traces
45mTracing: A journey to tactical insights
31mFrom Alerts to Insights: Performing Trace-Based Causation at Scale
25mProfiling Request Latency with Critical Path Analysis
24mAutomated Root Cause Analysis