Review Perf

Review Perf

Follow shared context and the findings contract. Own cost classification/measurement and runtime clock/randomness injection. Test-only utilities belong to review-tests, signal contracts to review-telemetry, and abstractions without cost claims to review-naming. Reuse telemetry/resilience evidence without duplicating the same allocation or per-call finding.

Procedure

1. Classify frequency: per-request, per-item, per-connection or startup-only. Only the first three earn optimization pressure. Off-path complexity to avoid allocations can itself be a finding.
2. Apply the cost/injection questions below.
3. Benchmark faster/slower claims, preferring the repository's existing harness. Never infer a regression by reading code; unmeasured runtime suspicions are conditional questions.

Specialist questions

• Allocation: For static text prefer Cow<'static, str>, HeaderName, HeaderValue, or common enum variants with Other(..) over String. Remove needless str -> Uri -> str, reflexive .clone() and to_string() on owned values.
• Frequency/dispatch: Resolve invariant flags when building the pipeline or service, not per request. Prefer hot-path static dispatch; measured type erasure at a stored pipeline's edge is legitimate, not grounds for generics for their own sake.
• Growth/contention: Check repeated work, quadratic scans, locks and per-call synchronization. Consider specializing single-entry caches/fast paths. Chunk or bound external-input allocations; never trust caller sizes.
• Determinism: Flag tokio::time::sleep, Instant::now() and SystemTime::now() where clock abstractions exist, and ad-hoc entropy where a seedable source exists. Injection is also a testability requirement.

Proof and coverage

Report benchmark point estimate, range and statistical significance: e.g. "0.21 ns (4%), not significant in Criterion, so insufficient to justify the reference." Prefer allocation/instruction-count regression guards over timing where supported.

Identify path frequency, avoidable work or hard-wired source, decisive evidence and measurements, concrete cost/testability impact, and a correction without unjustified complexity. Keep unmeasured runtime claims conditional; off-path or unmeasured findings are Non-blocking. Report performance neutrality rather than inventing micro-optimizations.

Coverage: paths classified, measurements and unmeasured areas.