egorfedorov/low-latency-systems

Low Latency Systems

Use this skill to turn latency incidents and regressions into measurable, reproducible fixes.

Workflow

1. Lock measurement context first.

• Capture workload, concurrency, payload sizes, warmup policy, and hardware/runtime settings.
• Keep baseline and current runs environment-compatible.

2. Decompose latency path.

• Split end-to-end latency into ingress, queue, compute, storage/network, and egress components.
• Prioritize tail-latency contributors over average-only improvements.

3. Apply targeted latency fixes.

• Reduce blocking, contention, and unbounded queues.
• Reduce allocations/serialization overhead in hot paths.
• Use batching, caching, and async boundaries only when measured beneficial.

4. Validate percentile regressions.

• Compare baseline vs current percentiles (p50, p95, p99, optional p999).
• Gate release on configured regression thresholds.

5. Produce sign-off output.

• Provide measured deltas, affected components/files, and residual risks.
• Include exact rerun commands for verification.

Commands

python3 scripts/compare_latency_runs.py \
  --baseline <baseline.json> \
  --current <current.json> \
  --threshold-pct 5

Treat non-zero exits as blocker regressions.

Output Contract

Return:

1. Latency Baseline: environment/workload assumptions.
2. Findings: percentile deltas and hotspot classes.
3. Optimization Plan: exact changes with expected impact.
4. Verification: rerun commands and regression gates.
5. Residual Risks: variance or unresolved tail spikes.

References

• references/workflow.md: detailed low-latency process.
• references/latency-playbook.md: bottleneck-to-fix mapping.
• references/signoff-template.md: concise sign-off format.

Execution Rules

• Prioritize tail latency (p95/p99) when evaluating user impact.
• Keep measurement setup stable across comparisons.
• Require before/after evidence for each claimed improvement.
• Escalate threshold breaches as blockers.