[mbreiser]

Summary

Two host-monitor fixes for the next package set, rig-nodes 0.3.1 (LAB-338). Both were found validating 0.3.0 in the webDisplayTools acceptance harness: typed on ww18, and the psutil path on the Mac.

1. Typed HostStats floats are quantized to power-of-two steps (host_sampler.quantize, applied by host_monitor.to_msg).

• Problem: float32 carries ~24 bits of sampling noise per value, which zstd cannot compress. Typed /rig/host was 6.7 % of a ww18 harness run's compressed bag (1.77 MB/h), against 4.8 % for the String JSON form. The harness gate is 5 %.
• Steps: 0..1 fractions and CPU cores to 1/1024 (0.1 %); everything else measured (rates, MB, MB/s, ms, GB, °C, W) to 1/64. Values n·2^-k are exact in float32.
• Unchanged: NaN (absent) stays NaN, ints and strings are untouched, the caller's sample is not mutated, and the JSON path is unchanged.
• Result on ww18: typed /rig/host 4.0–4.1 % of the compressed payload (1.02–1.05 MB/h).

2. Portable path (macOS/Linux, psutil): the whole process table every 5 s, the watched pids in between (HostSampler._ps_rows, proc_rescan_s, CLI --proc-rescan).

• Problem: every window read cpu_percent + memory_info for every process, so the monitor cost 6.7 % of a core on the lab Mac (18 CPUs; the harness gate is 2 %).
• Fix: the whole table is read every proc_rescan_s. In between, only the watched pids are read: each pinned group's processes, the last window's top-N rows, and the monitor itself.
• Behaviour:
  • A Process object keeps its cpu_percent baseline, so an unwatched process's value at a rescan averages the interval; a newly busy process appears at the next rescan.
  • The pid → group cache is kept against the full table, so Python command lines are not re-read each rescan.
  • Windows is unchanged: one NtQuerySystemInformation call per window.
• Result, portable path forced on ww18 (psutil on Windows, ~300 processes): monitor CPU mean 24.9 % → 7.3 % of a core, median 24.8 % → 1.35 %. The rescan windows peak at ~35 %. The Mac number is to be measured on this branch.

Test plan

• tests/test_host_sampler.py: 52/52 (+18), covering:
  • quantization: grid, nearest rounding, ≤ half-step error, NaN, ints/strings untouched, no mutation, float32-exact;
  • rescan: one full read per period, unwatched pids read only then, pinned pids every window, the watch set, rows equal to a full read.
• tests/test_pure.py: 40/40
• webDisplayTools acceptance harness on ww18 (Windows 11, 24 CPUs), typed environment, with this branch's source first on PYTHONPATH (candidate 037a5cc + rig: record the rig-nodes host monitor (/rig/host) in every run's bag webDisplayTools#282): exit 0, every check passes, /rig/host typed 4.0 % of the compressed payload, monitor CPU median 0.9 % of one core
• Mac (18 logical CPUs, portable path): pixi run test-host 52/52.

  • Standalone sampler, 20 windows, gpu=False, CPU ms per window by part:

    	median	mean	max	_processes	_core_busy	total
    installed 0.3.0	7.1 %	7.0 %	7.5 %	78.6 ms	2.5 ms	82.1 ms
    this branch	0.56 %	1.9 %	7.4 % (the 1-in-5 rescan windows)	18.7 ms	2.6 ms	22.7 ms

  • The process table was the whole macOS cost.

  • (A first harness run on the Mac seemed to show no change; it had loaded the installed 0.3.0, because pixi's ROS activation overwrites PYTHONPATH.)

🤖 Generated with Claude Code