Files
windmill/benchmarks/sim/oom_poller.ts
pyranota 74b662d8de feat(benchmarks): k8s sim mode + util-group dashboard + reliability fixes
Stand up a minikube-backed simulation subsystem for benching Windmill under
realistic multi-node load, with a per-bench measurement pipeline and a
dashboard renderer that consolidates throughput, queue depth, per-node CPU,
PG latency/conns, OOM events, and per-node CPU-util-vs-oversaturation into
one SVG report.

Sim infrastructure (sim/):
- k8s_provisioner: minikube up + heterogeneous node sizing from topology JSON
- helm_deploy: helm install Windmill with smoke.yaml + local.yaml overlays
- image_cache: pre-load required images so bench bringup is offline-safe
- toxiproxy_k8s: per-node toxiproxy DaemonSet for cross-node latency injection
- cpu_sampler_k8s: privileged DS reading per-cgroup cpu.stat at 10Hz, dual-
  writes to stdout AND a host-mounted log file (/var/log/wm-sim-cpu-sampler/
  sampler.tsv) so heavy benches no longer lose early samples to kubelet log
  rotation
- pg_logging: ALTER SYSTEM + SIGHUP to enable verbose PG logging without restart
- pgbadger: post-bench PG log analysis HTML report
- readiness: pre-bench cluster health check (samplers stable ≥30s, workers
  ready, PG responsive, queue empty, **deploy.status rollout-complete**) —
  the rollout-complete check catches mid-rolling-update fires that previously
  starved m04's sampler under cgroup_mutex contention

Per-bench JSONL pollers, started/finalized alongside the bench loop:
- pod_timeline: 1Hz workers-per-node Ready counts (used for the workers panel)
- oom_poller: live OOM event capture (kernel + kubelet evictions + cgroup)
- pg_latency_poller: 4Hz psql \\timing on SELECT 1 vs kubectl-exec roundtrip
- pg_conn_poller: 1Hz pg_stat_activity by state (active/idle/idle_in_xact)
- node_load_poller: 2Hz /proc/loadavg + /proc/stat procs_running per node

Dashboard renderer (sim/render_report.ts + graph.ts):
- Util group: one panel per node with translucent orange oversaturation area
  BEHIND solid blue CPU-util area, 100% reference line, phase-boundary verticals.
  cols:2 grid wraps after 2 panels per row.
- PG node tinted with [PG] flag in legend across the dashboard.
- Phase-boundary verticals + push-window shaded zones layered consistently.
- All x-axes switched from wall-clock HH:MM to relative seconds-from-bench-
  start. Shared origin sourced from meta.json's bench_start_ms so 0s on every
  panel = the same wall-clock moment (previously each chart picked its own
  earliest sample as origin, causing drift between panels).

Oversaturation metric, with explicit fallback:
- Primary: (procs_running - ncpu) / ncpu × 100 — true CPU run-queue pressure.
- Fallback to load1 when procs_running is missing (older reports).
- load1 overcounted previously because it includes uninterruptible D-state
  procs (PG backends in disk I/O, cgroup_mutex waits), inflating "saturation"
  by 5-10x under load.
- Pure helper extracted to sim/util_metrics.ts; 8 unit tests cover the
  procs_running > load1 preference, the clamp-at-zero, invalid-ncpu cases.

Sampler reliability:
- HostPath log file in addition to stdout so the bench's scp-based collector
  bypasses kubelet log rotation entirely.
- main.ts truncates the host log file on every node before pushers start
  (parallel ssh, best-effort) so it doesn't grow unbounded across runs.
- Collector falls back to kubectl-logs when scp fails for any node.

Workloads (workloads/):
- io_4phase: four-phase IO step (idle → 2.5s → 500ms → 150ms jobs)
- io_150ms_flood / io_300ms_flood / io_1s_flood / io_2s_flood: single-phase
  flood configs to isolate the worker-host CFS context-switch storm vs PG
  contention regime
- burst, ops_day, cpu_*, etc. for other scenarios

Tests:
- sim/util_metrics_test.ts — 8 cases for computeOversatPct
- sim/util_panel_snapshot_test.ts — 5 assertions guarding util-panel SVG
  invariants (orange behind blue, 100% ref line, relative-time ticks NOT
  wall-clock, phase-boundary verticals, shared-origin override)

Helm values:
- sim/values/smoke.yaml — bench-tuned: workers w/ no CPU limit & low mem
  request, PG w/ 3-core request + wm-critical priorityClass + oomImmune +
  maxConnections, app w/ wm-critical + oomImmune + no resource limits.
- sim/values/local.example.yaml — template for the gitignored local.yaml
  that carries the EE license key.
- Depends on the wm-critical PriorityClass + oomImmune + maxConnections
  knobs landing in windmill-helm-charts (separate PR).

graph.ts additions:
- areaFills param: ordered list of per-kind translucent area fills drawn
  before lines, used by the util panel for orange-behind-blue layering
- lineColorOverrides: pin per-kind line colors so oversaturation reliably
  renders orange regardless of d3 ordinal-color insertion order
- highlightKindToken: substring-match flag for the PG-node tint in Node CPU
- xRelativeOriginMs: shared bench-start origin for the relative-time x-axis
- DataPointMulti is now exported for downstream tests

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-06-08 11:43:47 +02:00

100 lines
3.6 KiB
TypeScript

// In-bench OOMKilled poller. The end-of-bench scan in `oom_events.ts` reads
// `lastState.terminated` from every pod, but that field only holds the MOST
// RECENT termination — a pod that OOMKilled mid-bench and then recovered into
// Running state by end-of-bench shows no OOMKill at all in the final scan,
// silently zeroing out the L2 cgroup OOM panel.
//
// This poller closes the gap: every N seconds it lists all pods, watches for
// `restartCount` increases on any container whose `lastState.terminated.reason`
// is "OOMKilled", and appends a JSONL row per newly-seen OOMKill. End-of-bench
// `collectOomEvents` reads this JSONL and merges it with the final scan so the
// L2 cgroup OOM panel reflects every kill that fired during the window.
//
// Dedupe key: `<pod>|<container>|<finishedAt>` — restartCount can climb across
// multiple polls before the next OOMKill, but finishedAt is unique per kill.
import { MinikubeProvisioner } from "./k8s_provisioner.ts";
export type OomPoller = {
// Mutate this to false to stop the loop. The polling promise resolves when
// the in-flight kubectl call finishes after the flip.
cont: { value: boolean };
done: Promise<void>;
};
type ContainerStatus = {
name?: string;
restartCount?: number;
lastState?: {
terminated?: {
reason?: string;
finishedAt?: string;
};
};
};
type PodSnapshot = {
metadata?: { name?: string };
spec?: { nodeName?: string };
status?: { containerStatuses?: ContainerStatus[] };
};
export function startOomPoller(
prov: MinikubeProvisioner,
outPath: string,
opts: { intervalMs?: number } = {},
): OomPoller {
const intervalMs = opts.intervalMs ?? 2000;
const cont = { value: true };
const f = Deno.openSync(outPath, { write: true, create: true, truncate: true });
const enc = new TextEncoder();
// Dedupe: a kill is uniquely identified by where + when it finished.
const seen = new Set<string>();
const done = (async () => {
while (cont.value) {
const startMs = Date.now();
try {
const res = await prov.kubectl(["get", "pods", "-A", "-o", "json"]);
if (res.code === 0 && res.stdout) {
let parsed: { items?: PodSnapshot[] } = {};
try { parsed = JSON.parse(res.stdout); } catch { /* transient — retry next tick */ }
for (const p of parsed.items ?? []) {
const podName = p.metadata?.name;
const node = p.spec?.nodeName ?? "unknown";
if (!podName) continue;
for (const cs of p.status?.containerStatuses ?? []) {
const term = cs.lastState?.terminated;
if (term?.reason !== "OOMKilled") continue;
const finishedAt = term.finishedAt;
if (!finishedAt) continue;
const key = `${podName}|${cs.name ?? "?"}|${finishedAt}`;
if (seen.has(key)) continue;
seen.add(key);
const ts = Date.parse(finishedAt);
const row = JSON.stringify({
ts_ms: Number.isFinite(ts) ? ts : startMs,
source: "cgroup",
victim: podName,
container: cs.name,
node,
});
f.writeSync(enc.encode(row + "\n"));
}
}
}
} catch (_e) {
// Transient kubectl errors tolerated — bench should not die because
// one poll failed. Next iteration retries.
}
const elapsed = Date.now() - startMs;
if (cont.value && elapsed < intervalMs) {
await new Promise((r) => setTimeout(r, intervalMs - elapsed));
}
}
f.close();
})();
return { cont, done };
}