Performance

Measure your application before changing concurrency, caching or pipeline order. This RC does not establish a universal latency/throughput target.

Measure the relevant boundary

Use representative payloads and concurrency. Record p50/p95/p99 latency, throughput, errors, CPU/memory and external-service timings. Separate cold start, warm traffic, cache hits/misses and overload. Include tenant-specific workloads where relevant.

Use public pipeline diagnostics, metrics and tracing. The package does not export a TelemetryService singleton or trackPerformance helper under those names.

Optimize without changing security semantics

• Keep authentication, authorization, validation and rate limits in their required order. An optimizer suggestion is not proof of semantic equivalence.
• Reuse properly scoped connections and batch bounded independent work.
• Use cache policies only for data safe to share. Separate tenant/host variants and invalidate on relevant writes.
• Bound job concurrency and apply backpressure instead of unbounded promises.
• Use browser entry points and inspect bundle output for Node dependencies.
• Configure collector/export intervals and avoid high-cardinality labels.

Use public composition helpers rather than unexported middleware registries or SSR/ISR subpaths. Cancellation, timeout responses and underlying operation cancellation are distinct: confirm that timed-out work cannot continue causing unwanted side effects.

See deployment and testing.