Optimization Pattern — Workflow

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

---

End-to-end flow

Step procedure

Step	Actor	Action	Exit criteria
1	Human / Perf owner	Freeze metric, min effect, guardrails, abort thresholds	Written success/fail definition
2	Workload Curator	Check in load profile + fixtures	Versioned workload ID
3	Benchmark Runner	Baseline: warm-up, n samples, p50/p95/p99 + CI	Baseline artifact stored
4	Profiler	Attribute CPU/IO/alloc/lock/token cost on critical path	Dominant frame/span named
5	Optimization Agent	One hypothesis; reject multi-change bundles	Single mechanism stated
6	Coder	Implement behind flag only	Diff limited to hypothesis + tests
7	Benchmark Runner	Candidate vs control under identical load	Effect size ± CI; oracle green
8	Canary Observer	Progressive exposure; watch abort metrics	Hold times met or abort
9	Human	Promote or kill; document capacity note	Flag state + headroom recorded

Abort criteria (must be pre-declared)

Abort canary and disable flag if any hold for the configured window (e.g. 10 minutes):

• Target metric regresses vs baseline beyond noise band
• Guardrail: error rate > baseline + absolute delta (e.g. +0.1%)
• Guardrail: p99 > SLO or saturation (CPU/mem/queue) > soft limit
• Oracle failures in synthetic probes
• Neighbor-tenant latency unfairness alert fires

Parallelism rule

Never run two optimization hypotheses on the same critical path concurrently. Confounded A/B results are not evidence. Serialize changes; keep one flag per hypothesis.

Deviation rule

If implementation discovers the bottleneck moved or the change needs a second mechanism: stop, re-profile, open a new hypothesis. Do not silently widen scope behind the same flag.