19 Releases 150 Tags

RSS Feed

• v0.10.10-beta.1 2a17dfc370

  Compare

  perf(metrics): scan the OS once per tick; cache every static fact the collectors used to re-read

  LastSkywalker released this 2026-09-08 21:09:46 +03:00 | 64 commits to main since this release

  The metrics tick collected every module twice on persist cycles (CollectAndSave
  per module, then CollectAllCurrent for the live fan-out) and the collectors
  re-derived static facts on every pass. Measured under strace on Linux: 208
  traced syscalls per tick before, 26 after, with the same payloads.

  Pipeline

  • One system.Snapshot per tick: history asks the system service for a single
    parallel pass and hands the same value to the DB writer (persist ticks), the
    SSE broker, the cluster metric stream and the Prometheus exporter. Concurrent
    CollectSnapshot calls coalesce (ticker + on-demand GET /metrics/current), so
    two passes can no longer interleave inside the network/CPU rate baselines.
  • GET /metrics/current serves the latest snapshot while younger than one tick.
  • The SSE body is marshaled once from the struct (no marshal→unmarshal→marshal
    to inject collecting_host_id); the host row comes from the last registration
    instead of two GetCurrentHost round-trips (8 SELECTs) per tick.
  • WireGate hashes the docker inventory off the struct; on an idle tick the
    docker payload is never marshaled for the wire at all.
  • Prometheus exports the tick's snapshot instead of re-sampling every collector
    (which also reset the SSE stream's network rate baseline).

  Collectors

  • cpu: cpu.Info (1 + N-cpus file opens, only [0] used) cached, 10-min refresh.
  • memory: host paths resolved once; the sysfs memory-block walk (one file per
    128 MB block) cached 10 min; swap reuses the meminfo parse.
  • disk: mount table cached 60 s; pseudo filesystems skipped BEFORE statfs
    (node_exporter's default exclude lists); statfs behind a watchdog with a
    stuck-mount set so a dead network mount can't wedge the tick; Linux reads
    /proc/diskstats directly and refreshes serial/label (two udev scans per
    device per call in gopsutil) every 10 min.
  • network + hosts: addresses/MACs/primary route from a shared 60-s topology
    cache (hostnet.CurrentTopology) with early refresh when a new interface name
    appears; gopsutil's net.Interfaces was one netlink dump per interface per tick.
    Counter rollback (interface re-created) re-baselines instead of wrapping.
  • hosts: OS identity (host.Info forks lsb_release and lists /proc) cached an
    hour; boot time cached via gopsutil's opt-in.
  • sensors: new internal/platform/hwsensors — one cached, time-bounded read per
    tick shared by the cpu collector and the sensors endpoint (the cpu collector
    also retried the identical call on a warning). Linux discovers the hwmon file
    set once and re-reads only temp*_input with a fixed-buffer read(2).

  Docker

  • Inspect details cached per container (key = state|created, 5-min resync):
    one ContainerList + one one-shot stats per running container per tick
    instead of N extra inspects for values fixed per container lifecycle.
  • Image-check targets built only when the hourly check is due; gateway route
    source read at most every 30 s; debug dump loops gated on the log level.

  Verified on the native macOS binary, a Docker instance on the rebuilt image
  (host-mounted /proc,/sys,/ paths) and the live nsdemo stack: one collection per
  10 s in the logs, REST/SSE payloads unchanged, the machines/stats/containers/
  application pages render and update live over SSE.

  Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com

  Downloads

  • Source Code (ZIP)
  • Source Code (TAR.GZ)