Worker launch-latency alternatives — 2026-08-10

Verdict

INCOMPLETE. The local packet executes fak’s summary and a raw-event/no-summary baseline only. Prometheus, OpenTelemetry, and Datadog have zero measurements until real ingestion and queries run; #6137 tracks those witnesses.

Same-workload contract

Every arm consumes the same six dispatch/heartbeat observations spanning fixed [1,2,5,10,30] second edges, including one heartbeat preceding dispatch. Every backend must use the same lower-closed bucket convention, nearest-rank p50/p95 convention, and independent oracle. Full runs report bucket/quantile quality, dropped observations, ingestion/query latency, CPU/RSS/network/storage, and total cost.

Arm	Class	Local availability	Honest result
fak native launch-latency summary	native	yes	correct six-sample histogram, p50=2s, p95=30s, one negative
raw launch events without summary	tuned no-feature baseline	yes	incorrect: no buckets or percentiles
Prometheus histogram	external	no	zero measurements
OpenTelemetry metrics	external	no	zero measurements
Datadog distribution metric	external	no	zero measurements

No equivalent first-class fak backend integration was identified, so none is fabricated. Adapters and mocks are not backend witnesses.

Local native witness

go test ./internal/launchlatency -bench BenchmarkLaunchLatencySummary -benchmem -run '^$' -count=5

Windows/amd64, AMD Ryzen 9 9950X: 159.0, 159.8, 203.8, 209.2, 198.7 ns/op. Median: 198.7 ns/summary, 528 B/op, 6 allocs/op. This measures in-process folding only, not telemetry ingestion/query latency or cost.

Reproduce

go test ./internal/launchlatency -run TestCompareLocalKeepsTelemetryAlternativesExplicit
go test ./internal/launchlatency -bench BenchmarkLaunchLatencySummary -benchmem -run '^$' -count=5
go test ./internal/nativebench