Skip to content

bench-targets.md

Latest commit

 

History

History
73 lines (60 loc) · 3.99 KB

bench-targets.md

File metadata and controls

73 lines (60 loc) · 3.99 KB
Raw

Performance Target Rubric v2.4

Fitz keeps its machine-readable performance targets in config/perf_targets.json. This document is the human-readable mirror of the 14 primary Tier 4 throughput rows in config/bench_release_ids.txt.

Baseline

These current values come from the clean default-profile run generated on 2026-08-07:

  • CPU: 12th Gen Intel Core i9-12900HK, 20 logical cores
  • RAM: 62 GiB
  • OS: Linux x86_64
  • Network: loopback, without TLS
  • Build: release
  • Release storage: memory-backed KV and Stream wire lifecycle gates
  • Storage coverage: local-disk rows remain deep storage_characterization results
  • Primary comparison metric: mean microseconds per completed operation

Throughput records are normalized to mean_us as 1_000_000 / operations_per_second. Lower values are better. The derived minimum operations per second columns are retained for quick scanning, while mean_us remains canonical.

The KV and Stream thresholds were reset around their newly promoted memory-backed rows: operational maximum is 20% above the current mean and stretch maximum is 10% above it. The other domain thresholds remain the previously reviewed operational goals, with only their current measurements refreshed.

Release Regression Gates

All rows use target_class = regression_gate, budget_group = tier4_integration, and hard gating.

domain suite scenario layer current us operational max us min ops/sec stretch max us min ops/sec
kv tier4-kv-gate begin_put_commit_lifecycle tcp 83.149 99.779 10,022 91.464 10,933
kv tier4-kv-gate begin_put_commit_lifecycle websocket 81.590 97.908 10,213 89.749 11,142
lease tier4-lease-gate acquire_release_lifecycle tcp 49.409 73.740 13,018 67.590 14,645
lease tier4-lease-gate acquire_release_lifecycle websocket 48.758 80.040 11,992 73.370 13,492
notice tier4-notice-publish delivery_confirmed_publish tcp 28.813 47.000 20,421 43.090 22,974
notice tier4-notice-publish delivery_confirmed_publish websocket 24.721 45.770 20,973 41.950 23,595
queue tier4-queue-gate enqueue_reserve_ack_lifecycle tcp 94.530 308.940 3,107 283.190 3,495
queue tier4-queue-gate enqueue_reserve_ack_lifecycle websocket 99.158 216.570 4,432 198.520 4,986
rpc tier4-rpc-roundtrip request_response_roundtrip tcp 50.952 60.300 15,919 55.270 17,909
rpc tier4-rpc-roundtrip request_response_roundtrip websocket 50.321 61.480 15,613 56.360 17,565
schedule tier4-schedule-lifecycle create_fire_ack_lifecycle tcp 69.762 130.690 7,345 119.800 8,263
schedule tier4-schedule-lifecycle create_fire_ack_lifecycle websocket 71.942 132.570 7,241 121.520 8,146
stream tier4-stream-gate sync_write_lifecycle tcp 153.630 184.356 5,424 168.993 5,917
stream tier4-stream-gate sync_write_lifecycle websocket 150.158 180.190 5,549 165.174 6,054

Interpretation

  • Release acceptance uses primary throughput records only. Paired latency records remain diagnostic because their per-operation distributions are intentionally noisier than the duration-window throughput signal.
  • Local-disk KV, Stream, Queue, and Schedule rows still appear in the full benchmark report. A noisy storage characterization row is a review signal, not a release-manifest failure.
  • Add a workload to this rubric only after it is present in a clean full baseline, stable under the default profile, and intentionally promoted to the release manifest.

Document History

Date Version Changes
2026-08-07 2.4 Rebuilt the rubric from the clean 14-row release manifest, removed stale workload expectations, and promoted memory-backed KV and Stream transport gates.
2026-07-07 2.3 Previous target rubric.