Skip to content

Observability overhead gate

  • Overall result: PASS
  • Observed: 2026-07-11 on the Primary Node Windows workstation, using the complete default Windows, WSL/Docker, NVIDIA, container, service-health, retention, and dashboard-sampler path.
  • Raw artifact: anvil-evidence://observability/observability-overhead.json.gz
  • SHA-256: 98da7a3426493e6f9aef92c7e663ab8a7ec2a596b0824002a1f0f841022c983e
  • Compressed size: 596 bytes

Normal profile

  • Result: PASS
  • CPU average / peak: 0.3051% / 3.1323%
  • RSS average / peak: 37640520 / 38699008 bytes
  • Disk writes / GPU allocation: 0 / 0 bytes
  • Throughput / latency change: 0.3529% / -0.3516%
  • Child-process peak CPU / RSS: 9.7487% of host capacity / 99,512,320 bytes
  • Docker / WSL command invocations: 10 / 0
  • Failures: none

Benchmark profile

  • Result: PASS
  • CPU average / peak: 0.1513% / 0.2716%
  • RSS average / peak: 38563185 / 39501824 bytes
  • Disk writes / GPU allocation: 0 / 0 bytes
  • Throughput / latency change: 0.3529% / -0.3516%
  • Child-process peak CPU / RSS: 9.8206% of host capacity / 99,028,992 bytes
  • Docker / WSL command invocations: 12 / 0
  • Failures: none

Method and interpretation

The normal sampler used the approved 2-second core and 5-second costly cadences. The benchmark sampler used 1-second core and 2-second costly cadences. The harness sampled the Python process at 250 ms, measured process CPU against total logical host capacity, read Windows working-set and process I/O counters, queried NVIDIA compute-process allocation for the dashboard PID, and watched collector child processes at 5 ms for peak CPU and RSS. The raw artifact contains the exact unrounded results and profile limits.

The collection-effect check bracketed a six-second collection-on control with two three-second collection-off controls and compared median latency and its reciprocal throughput across 295 collection-on samples. It is an overhead control workload, not a model-quality or model-throughput benchmark, and does not change any model recommendation in docs/BENCHMARKS.md.

The benchmark-mode costly frame reported 87 degraded samples. Inspection tied these to explicitly unavailable per-container GPU attribution, not failed collection; the core frame was healthy. Docker and WSL cost is reported as collector command invocations because those are the external API boundaries in the stdlib collector path. Zero capture writes refers to the dashboard Python process and its in-memory retention path; raw evidence was compressed only after measurement ended.