Triage Latency Regression

When to use

Use when p95/p99 latency increases and you need a fast root-cause hypothesis.

Preconditions

• You have access to metrics, logs, and (ideally) traces.

Procedure

1. Confirm scope: which endpoints/operations, which regions, which customers.
2. Compare before/after: deployment markers, config changes, dependency status.
3. Decompose: queueing vs compute vs dependency latency.
4. Validate saturation: CPU, memory, GC, thread pools, connection pools.
5. Correlate errors/timeouts with latency.

Decision points

• Strong correlation with a deploy: consider rollback.
• Dependency latency increase: mitigate with timeouts, caching, failover.
• Saturation increase: scale, shed load, or reduce work.

Verification

• Latency returns to baseline and stays stable across multiple windows.
• Error rate does not regress.

Rollback / undo

• Roll back the most recent change correlated with the regression.
• Revert performance tuning if it degrades correctness.

Escalation

• Escalate to dependency owners if external system is the bottleneck.
• Escalate to service owners if code changes are required.

Examples

Example queries to adapt:

Compare p95 latency now vs 1h ago; split by route, status code, and upstream.