Debug a Latency Regression
After a deployment, p99 latency triples while p50 stays stable. Outline a safe investigation and mitigation using metrics, traces, logs, and controlled rollback.
Examples
Compare approaches
Baseline approach
Begin by explaining the core mechanism, stating assumptions, and walking through one concrete example.
Time: Depends on design · Space: Depends on design
// Write hypotheses, evidence, and rollback criteria.
Strong answer
Start with impact and deployment correlation. Segment by endpoint, region, and dependency. Inspect slow traces and queueing, mitigate with rollback or traffic controls, and preserve evidence.
Time: Discuss operation costs · Space: Discuss retained state
// Write hypotheses, evidence, and rollback criteria.
Common traps
- Averages hide tail latency. Avoid changing many variables at once.
- State assumptions and justify trade-offs rather than memorizing a single answer.