Isolated Routing Scaffold Benchmark

May 11, 2026 · View on GitHub

WisePick Decision API · v0.1.3 · harness output (tests/stress_test.py)

  • Concurrent segment wall time (perf_counter): 11.8092 s
  • Worker pool max_workers: 512

Benchmark scope

This run is an isolated routing scaffold benchmark. It measures in-process work for run_decision only.

IncludedExcluded (stubbed or bypassed)
run_decision CPU path: capability extraction, scoring, winner selectionHTTP stack, FastAPI, serialization over the wire
Per-request elapsed time via perf_counter (ms)Live PostgreSQL I/O; real decisions inserts
Concurrent load via thread pool (GIL-visible)Network calls to Yantrik
Fixed intent → baseline provider / capability_id checksEnd-to-end API latency under real DB pools / replication lag

Do not treat these figures as production SLOs. They bound the routing scaffold under this harness; deployed latency adds persistence, optional plugins, and runtime overhead.

Summary

MetricValue
Total requests20,000
Worker pool size (threads)512
Throughput (RPS)1,693.60
Mean latency (ms), concurrent load2.2578
P50 latency (ms), concurrent load0.5249
P95 latency (ms), concurrent load0.8913
P99 latency (ms), concurrent load55.1734
Success rate (%)100.00
Routing accuracy (%)100.00
Concurrent mean under 1 ms goal metNo
Single-threaded sample count10,000
Mean latency (ms), single-threaded0.6033
P50 latency (ms), single-threaded0.5594
P95 latency (ms), single-threaded0.8673
P99 latency (ms), single-threaded1.0430
Single-threaded mean under 1 ms (scaffold)Yes
Gate: single-threaded P95 < 1 msYes
Gate: concurrent segment RPS ≥ 1,000Yes
Errors0
Expected provider / capabilitydeepl_translate / translation

Measurement

  • Throughput (RPS): total_requests / wall_seconds for the concurrent segment (wall-clock for the whole phase, not per-thread service rate).
  • Routing accuracy: Share of responses matching the warmed baseline for the fixed task string.
  • Percentiles (P50 / P95 / P99): Computed on sorted per-request durations (ms); same linear-interpolation quantile function for all listed percentiles.
  • Concurrent vs sequential: Concurrent samples include scheduler and GIL contention; sequential samples approximate single-threaded scaffold cost without that contention.

Outcome (this run)

Single-threaded path: mean 0.6033 ms, P95 0.8673 ms (below 1 ms). Concurrent segment: 1,693.60 RPS over the timed window; P50 / P95 / P99 0.5249 / 0.8913 / 55.1734 ms under load. Internal gates recorded in the summary table passed for P95 and RPS; concurrent mean remains above 1 ms, consistent with multi-thread scheduling on this profile.