Optimization Pattern — Problem

Pattern: Optimization
Component: problem.md
Version: 1.1 | Updated: 2026-07-16

---

Statement

Performance and cost work without a frozen baseline, attributed bottleneck, and measured effect size produces changes that look faster in isolation and fail (or regress) under production distributions — while burning sprint capacity.

Measurable symptoms

Symptom	How you detect it
No baseline	PR description has no before percentiles under a named workload
Wrong path	Flamegraph / trace shows <15% of critical-path time in the changed code
Tail silence	Only mean or “feels faster” reported; p95/p99 absent
No kill switch	Optimization ships without a feature flag or config revert
Oracle gap	Benchmark never asserts business/correctness invariants

Root cause

Engineers and agents optimize what is easy to measure locally (microbenchmarks, single-request traces) instead of the metric that defines success under representative load, including variance and shifted bottlenecks.

Non-goals of this pattern

• Not a substitute for fixing functional bugs
• Not a license for speculative rewrites (“rewrite in Rust”) without attribution
• Not capacity planning theater — must produce numbers that capacity owners can use
• Not incident containment (hotfix first)

Related failure modes

See failures.md for microbenchmark theater, mean-vs-tail traps, and correctness collateral.