Cache reuse-divergence detection alternatives — 2026-08-10

Verdict

INCOMPLETE. The local packet executes fak’s cross-provenance reuse-divergence fold and a raw-counter/no-detector baseline. Prometheus, OpenTelemetry, Prometheus rules, and Datadog retain zero measurements until real ingestion, storage, queries, and alert paths process the common trace; #6155 tracks those witnesses.

Same-workload contract

Every arm receives three cache records at the same 0.15 absolute-ratio tolerance: one stable witnessed/observed pair (0.80 versus 0.75), one divergent pair (0.80 versus 0.30), and one witnessed-only record. The oracle requires two comparable records, one single-class record, and exactly one alert naming the divergent source. Complete runs report true/false/missed alerts, ingest and alert latency, records/second, CPU/RSS/network/storage, and total cost.

Arm	Class	Local availability	Honest result
fak native reuse-divergence fold	native	yes	exact one-alert oracle
raw reuse counters without divergence detection	tuned baseline	yes	stores the axes but emits no alert
fak + Prometheus	first-class integration	no	real scrape/rule/alert path required
fak + OpenTelemetry	first-class integration	no	real export/collector path required
Prometheus recording and alerting rules	external	no	zero measurements
Datadog anomaly monitor	external	no	zero measurements

Adapters and local query-shaped functions are not monitoring-system witnesses.

Local native witness

go test ./internal/cachewitness -bench BenchmarkFoldReuseDivergence -benchmem -run '^$' -count=5

Windows/amd64, AMD Ryzen 9 9950X: 556.0, 542.2, 615.7, 589.4, 593.4 ns/op. Median: 589.4 ns/three-record fold, 496 B/op, 9 allocs/op. This is in-process folding cost, not scrape-to-alert latency.

Reproduce

go test ./internal/cachewitness -run TestCompareLocalKeepsDivergenceAlternativesExplicit
go test ./internal/cachewitness -bench BenchmarkFoldReuseDivergence -benchmem -run '^$' -count=5
go test ./internal/nativebench