Reference
Benchmarking roadmap
Condensed measurement plan: Phase 2 stage swap, k6 chat path, Phase 3 gRPC/corpus gates, Phase 4/5 matrix.
Benchmarking roadmap
Condensed from the workspace audit companion. Full detail: ibex-harness-milestone-audit-and-benchmarking-roadmap.md §3 (parent of this repo).
Phase 2 — real stages + chat path
| Synthetic | Replace with |
|---|---|
stageAuth | authcache LRU hit |
stageRateLimit | packages/ratelimit |
stageDirectiveResolve | directive Redis hit |
stagePromptInject | prompt inject |
| Composite | mock provider + full middleware |
- k6 default:
POST /v1/chat/completions(token/agent env). Health optional viaK6_USE_HEALTH. - Pin
baseline.jsononly after a real 2.6.1 run. Streaming line required on 2.6.2.
Phase 3 — corpus / IVFFlat / cache
- Context assembly: gRPC-native load (ghz / k6 gRPC), p95 < 50ms @ 50 VU
- Search: p95 < 30ms @ 100k memories; IVFFlat confirmed via EXPLAIN
- Embedding cache hit > 70%; hot-cache in latency breakdown
- Corpus fixtures at 1k / 10k / 100k
Phase 4 / 5
- Phase 4: failover latency, Lua rate-limit p99, async queue depth/drops, per-adapter overhead vs baseline
- Phase 5: SLO burn, chaos (Redis/auth/provider), full-path load; Critical Path Grafana from MONITORING.md §10.2 (pulled forward to late Phase 2 / early Phase 3)
Edit on GitHub
Last updated on