writeonce/docs/stories/databasev2/03-wal-checkpoint.md

track	iteration	was_language_iteration	status	readiness	chain
databasev2	3	32	done	ready	6

databasev2 3 — WAL checkpoint: disk space reclamation and bounded replay

Moved 2026-08-26 from the language track, where this was iteration 32. Part of Story — databasev2: the database beyond RAM. Content unchanged by the move; its dependencies are restated in that track index.

Format: product/story-iteration-template. Part of Story — one language, one runtime, one database, one binary — the track this iteration was authored in before the 2026-08-26 move.

Inserted 2026-08-21 (stage-3 guarantee refinement found the hole): the WAL is append-only FOREVER — no checkpoint, no truncation exists in the engine or anywhere on the roadmap. Disk grows without bound and replay time grows with history, so restart cost rises with every write the program ever made. RAM reclamation already exists (deleted rows free their slot — 04-db-binding.md: "Ids are never reused; slots are"); this iteration is the DISK half. LAST in the concurrency chain: stage 3 → 22 → 31 → 24 → 23 → 32 — it wants 22's measured replay/restart numbers to justify its policy and must compose with 23's group-commit write path.

BRAINSTORMED 2026-08-28. Spec: 2026-08-28-wal-checkpoint-design.md · plan: 2026-08-28-wal-checkpoint.md (6 tasks). Read .dev/reference/postgresql for this — and the conclusion was that Postgres' design is unavailable to us, which is what makes the simpler one legitimate.

The design in one sentence: compact the log by rewriting it as one record per live row into a temp file, then rename it over the live WAL. Recovery is completely unchanged — boot still opens one file and replays it — and the crash criterion is satisfied by the filesystem rather than by code we must get right.

Why one file works here and not in Postgres. Postgres never compacts its WAL: its records are page deltas, so a compacted redo log is not a store, and it must keep heap files, a control file, a redo pointer and a second recovery source. Ours are full row images — apply_record implements UPDATE as remove-then-recreate — so a compacted log is a complete store. That one difference deletes the control file, the redo pointer, the cutoff offset and the separate process from the design.

Forks settled: no snapshot format (the compacted log is the snapshot); one source, not two; volume-only trigger as a ratio against the last compaction's own measured output, with an absolute floor — no timer, because Postgres' timer exists to bound loss from unflushed buffers and we have none; stop-the-world, with the pause measured against a stated budget rather than assumed acceptable.

The coupling that would otherwise be found late — and was found in time: compaction moves every record, so it invalidates every WAL offset iteration 2's resident: keys stores. The compactor rebuilds the offset map as it writes. Recorded here while iteration 2's storage half was still unimplemented; it landed 2026-08-29 and the obligation was discharged (f606fc9), including a worse failure this note did not predict — see the hazard section at the end.

Measured on master 2026-08-28, grounding the whole iteration: seed 20000 leaves a 986 614-byte log; 20 000 updates take it to 2 590 262 bytes with the same live rows (2.6× history for no data), and boot+verify on that store is 155 ms.

Progress — landed 2026-08-29

#	Task	State
1	wo_wal_compact — rewrite, fsync, rename, fsync parent, reopen	✅ 8ea510d
2	a stale compaction temp is removed at open	✅ 8bfbd4b
3	the trigger (pure decision + env knobs) and the ordering guard	✅ 6dbcb9a
4	kill -9 DURING compaction — 40 rounds, mutation-proven	✅ 9b283d5
5	measure space, boot and the stop-the-world pause	✅ d87f65a
6	closeout	✅ this change

Measured

	checkpointing off	checkpointing on
WAL used	1 962 358 B	907 094 B
boot	114 ms	64 ms

2.16× space reclaimed, 1.78× faster boot, stop-the-world pause 2 651 µs against a stated 50 ms budget. Full details, including the pause's scaling, are in perf-targets.md §7.

Two bugs the work found, both mine

Wiring only the drain left WO_SHARDS=1 never compacting — its log grew forever (536 KB where the multi-shard run held 446 KB), because a statement on the owner shard never enters that drain. Both write paths now check.

The dump was 8× slower than it needed to be, flushing through the committing path and so paying one fdatasync per 256 records for durability that is worthless before the rename. One final barrier took the pause from 107 649 µs to 13 212 µs on a 2 MB live set — ~22 MB/s to ~181 MB/s.

Acceptance Criteria

Met:

• Given an aged store, when it is compacted, then disk is reclaimed. ✅ 2.16× on the full campaign, asserted rather than merely recorded — the leg fails if the log is not smaller with checkpointing on.
• Given the same store, when it boots, then replay is bounded by the live set rather than by history. ✅ 114 → 64 ms.
• Given kill -9 at ANY instant during a checkpoint, when the process restarts, then recovery produces the same consistent store as if the checkpoint had never started, with no acknowledged write lost. ✅ 40 rounds per run, 10 consecutive clean runs, and proven to have teeth: against the design's rejected alternative (in-place rewrite instead of rename) the battery fails every run with the log destroyed.
• Given the iteration-22 replay numbers, then a before/after delta is recorded. ✅ perf-targets.md §7.
• Given writes arriving while a checkpoint runs, then the ack contract holds. ✅ compaction runs only where nothing is staged, asserted by a test that stages and requires refusal; wo_wal_compact also refuses as a backstop.

Outstanding:

• The resident: keys offset map. Discharged 2026-08-29 by iteration 2's task 5d (f606fc9). The obligation written at the compactor in wal.c did its job: the implementer hit it there. See the hazard section below for what it caught — and for the second, worse failure it did not predict.
• The pause is O(live rows). At ~181 MB/s a 1 GB live set implies ~5.5 s, past any interactive budget. Incremental or forked copying was deliberately not bought in advance; this is the number to buy it against.

Goals

• Disk space is reclaimed. A checkpoint writes the live store as a snapshot and truncates the WAL behind it; deleted rows and overwritten versions stop occupying disk forever.
• Replay is bounded. Startup replays snapshot + WAL tail, not the program's whole write history — restart time becomes a function of store size, not store age.
• Every existing guarantee holds byte-for-byte. Ack-after-durable, replay-whole-or-not-at-all, torn-tail drop, ids never reused — a checkpoint changes where bytes live, never what an ack means. A crash DURING checkpoint recovers from the previous snapshot + full tail: the old WAL is not truncated until the new snapshot is durable.

Acceptance Criteria (draft — the spec refines)

• Given a store with N rows after many writes and deletes, when a checkpoint completes, then disk usage reflects the live rows (plus the WAL tail), and a restart replays snapshot + tail to the byte-identical store.
• Given kill -9 at ANY instant during a checkpoint, when the process restarts, then recovery produces the same consistent store as if the checkpoint had never started — no acknowledged write lost, no partial snapshot ever read.
• Given the iteration-22 restart benchmark re-run after checkpoint lands, when replay time is measured on an aged store, then the bounded-replay improvement is recorded as a before/after delta. The "before" now EXISTS (databasev2 1, 2026-08-27): db-bench's replay leg measures ≈5.5 µs per WAL record, and — the number this iteration is actually about — 1.9× the boot cost for an identical live dataset once the same rows have been updated once each (20 000 rows: 110 ms at 20 000 records, 211 ms at 40 000). Boot cost tracks history, not data, which is exactly what a checkpoint collapses. Metrics: replay.inserts.*, replay.history.*, replay.history_penalty_x.
• Given writes arriving while a checkpoint runs (the DB actor serializes statements; the checkpoint must not stall them beyond the stated budget), when the mixed load completes, then every ack held its durability contract and the tail contains exactly the post-snapshot writes.

Out Of Scope

• MVCC / multi-version reads — the store is update-in-place RAM; "old versions" exist only as WAL history, which is exactly what truncation reclaims.
• Incremental/streaming backup, point-in-time recovery — a snapshot is a recovery artifact here, not a backup product.
• Cross-shard checkpoint coordination — the WAL is owner-shard-only (stage 3's rule); one shard, one checkpoint.
• Compression, dedup, tiering — measure first (22), add only what a number justifies.

Info

Forks the spec must settle:

1. Snapshot format — a row-image dump of the live store (simple, O(live rows)) vs a rewritten-compacted WAL (reuses replay machinery, O(live rows) too but stays in one format). Leaning: row-image dump in the WAL's existing record grammar, so replay needs no second decoder.
2. Trigger policy — size threshold (WAL bytes vs snapshot bytes ratio), boot-time compaction, explicit call, or some mix. Leaning: ratio threshold checked at commit, plus manual trigger for tests; decided against 22's numbers.
3. Write availability during checkpoint — stop-the-world dump (simplest; the DB actor just runs one long "statement") vs fork-and-dump vs incremental copy. Leaning: measure the stop-the-world pause on the 1M-row store first (22); complexity only if the pause breaks a stated budget.
4. Composition with 23 — the snapshot's durability barrier rides the same per-shard ring (WRITE+FSYNC chain, then the truncate); ordering vs in-flight group commits must be stated normatively in 04-db-binding.md's WAL section.

Proposed Solution

Brainstorm → spec → plan after 23 lands (the write path it composes with). Correction (2026-08-27): this said "using 22's aged-store replay numbers as the policy input", but iteration 22 produced no such numbers — it proved restart correctness and never timed it, and bench/baseline.json carried zero replay metrics until databasev2 1 added them. The policy input is the replay leg's ≈5.5 µs/record and its 1.9× history penalty. Extend 04-db-binding.md's WAL section with the snapshot format the way the record grammar is documented today.

Hazard: compaction invalidates every resident: keys offset

Surfaced while refining this iteration and recorded here so it is not rediscovered late. Iteration 2's resident: keys stores a WAL byte offset per row and reads the row back with pread at that offset. Compaction — whichever of the two shapes below wins — rewrites the log and moves every record, so every stored offset becomes wrong. Not stale-but-readable: pointing at an arbitrary byte in a rewritten file, which is a correctness fault, not a performance one.

So the two iterations are coupled and the coupling has to be designed, not discovered: either compaction rebuilds the offset map as it rewrites (it knows both addresses, so this is the cheap direction), or the snapshot persists the map and compaction is forbidden while any resident: keys table is live. The first is almost certainly right, but it means compaction cannot be written as a pure file operation that ignores in-memory table state.

Settled 2026-08-29 — and the hazard was only half the danger

The first shape was implemented, in iteration 2's task 5d (f606fc9). Compaction re-points each row as it writes it, using a value-only map update that cannot rehash, so a walk in progress stays valid and no per-row buffer of new offsets is needed. Compaction is therefore not a pure file operation, exactly as predicted above.

What this section did not predict is the failure that would actually have struck first. It described stored offsets going stale — a pointer into a rewritten file. But the compactor walked the slab bitmap, and a keys-resident row holds no bitmap bit: its slot returns to the free list when the payload is dropped. Every such row would therefore have been omitted from the new log altogether. That is silent data loss, not a bad pointer, and rebuilding offsets would never have caught it — the rows would simply have been gone.

Both failure modes are now pinned by test_keys_resident_survives_compaction, which rewrites rows in hash order so the offsets genuinely move; a map left un-repointed lands on another row and fails the identity check rather than passing by luck.

A failure after any row has been re-pointed is fatal by design: the map would name offsets inside a temp file that the failure path unlinks, and the intact original log replays correctly, so stopping is strictly better than serving wrong rows.