Bounded cache-snapshot alternatives — 2026-08-10

Verdict

INCOMPLETE. The local packet executes fak’s bounded fsynced JSONL snapshot/replay and an unbounded append-only JSONL baseline. Prometheus, OpenTelemetry, SQLite, Prometheus TSDB, and ClickHouse retain zero measurements until their real stores execute the common retention workload; #6165 tracks those witnesses.

Same-workload contract

Every arm receives five ordered provider-cache turns representing 1,500 input tokens under a three-row retention requirement. The native oracle requires the latest three rows in order, two dropped rows, successful replay, and non-empty durable storage. The append-only baseline preserves all five rows and therefore fails the bounded-retention oracle.

The native implementation truncates, writes, flushes, and fsyncs the target. It is not atomic replacement; this packet does not claim otherwise. Complete store runs must additionally measure malformed/partial tails and crash-loss behavior, write/read latency and throughput, CPU/RSS/network/storage, represented tokens, and total cost.

Arm	Class	Local availability
fak native bounded fsynced JSONL snapshot	native	yes
unbounded append-only JSONL	tuned baseline	yes
fak + Prometheus	first-class integration	no
fak + OpenTelemetry	first-class integration	no
SQLite WAL	external	no
Prometheus TSDB	external	no
ClickHouse	external	no

Adapters and in-memory substitutes are not storage-system witnesses.

Local native witness

go test ./internal/vcachesnapshot -bench BenchmarkWriteReadBoundedSnapshot -benchmem -run '^$' -count=5

Windows/amd64, AMD Ryzen 9 9950X: 1,997,710; 1,962,634; 1,961,638; 1,966,473; 1,975,365 ns/op. Median: 1,966,473 ns/write+read cycle, 74,207 B/op, 39 allocs/op. This includes local fsync, not remote-store ingestion.

Reproduce

go test ./internal/vcachesnapshot -run TestCompareLocalKeepsSnapshotAlternativesExplicit
go test ./internal/vcachesnapshot -bench BenchmarkWriteReadBoundedSnapshot -benchmem -run '^$' -count=5
go test ./internal/nativebench