Uber Automates Dependency Analysis Across Its Service Mesh

Uber built an automated system to map service dependencies and classify how downstream failures propagate across its service mesh. RPC middleware in YARPC correlates caller and callee failures, labeling dependencies as fail-close, fail-open, or unknown.

• Dependencies are classified by failure correlation probability: ≥0.8 fail-close, ≤0.2 fail-open
• The system has indexed most fail-close anomalies across Uber's service mesh
• Distributed tracing was rejected: at 0.01% sampling, one failure takes about 5.5 hours to capture
• Requests route through the Muttley proxy; analysis is built into the YARPC framework for Go and Java