Plan for the approved spec. Code-free per the repo convention
(docs/plan/discarded.md:54); the executor writes the code.
- T1 wo_wal_compact: walk live rows via the bitmap, append one INSERT
each through the EXISTING append path, fsync, rename over the live log,
fsync the parent dir, reopen the descriptor. Test asserts BOTH that the
log shrank AND that a replay reproduces the same rows/ids/values —
shorter alone is worthless, a truncating bug also passes that
- T2 a stale temp file is removed at open and never read. The test uses
PLAUSIBLE records, not garbage: garbage would be rejected anyway and
would prove nothing
- T3 the trigger as a PURE decision (used bytes, last compaction's
measured output, floor) so it is unit-testable without a store; env
knobs for floor and ratio, which is what makes the policy testable at
all. No timer, with the reason. The check is called only where nothing
is staged, asserted by a test that stages and expects deferral
- T4 kill -9 DURING compaction, extending the existing fork-based crash
battery. Asserts the PROPERTY — the store equals the pre- or the
post-compaction content, never a mixture, and every acked id survives.
Run repeatedly and state the count: it is a race, one green run proves
little
- T5 measure space reclaimed, boot before/after, and the stop-the-world
PAUSE against a stated budget. If the pause exceeds it, stop and report
— the alternatives are bought against that number, not before it
- T6 closeout, including the normative ordering rule in 04-db-binding.md
Constraints carried from the spec into every task:
- recovery must NOT change; a task editing the replay path should stop
- the dump must FLUSH PERIODICALLY. stage() grows the staging buffer by
doubling, so dumping a whole store through one buffer would hold the
entire store in RAM — the unbounded growth databasev2 1 identified as
how this engine dies
- a FAILED compaction is a missed optimisation, not a durability event,
so it must not take databasev2 4's fatal path
- gate tolerances must not be waived wholesale (part A's T4 made that
mistake), and the baseline is full-mode — writing a quick-mode baseline
over it is a regression part A also made
Deliberately NOT a task: rebuilding the `resident: keys` offset map. It
cannot be implemented against a feature that does not exist yet, so T6
records it as an obligation at the compactor and in the story instead of
a stub nobody can test.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
databasev2 3, chain 6. Brainstormed 2026-08-28 after databasev2 4 part A
landed.
Design: compact the log by rewriting it as one record per live row into a
temp file, fsync, rename over the live WAL, fsync the parent dir, reopen.
Recovery is COMPLETELY UNCHANGED — boot still opens one file and replays
it — and the crash criterion ("the same store as if the checkpoint had
never started") is satisfied by rename, not by code we must get right.
Read .dev/reference/postgresql for this. The finding is that PG's design
is UNAVAILABLE to us, which is what makes the simpler option legitimate:
- PG never compacts its WAL; segments before the redo point are recycled
by rename or unlinked. Its records are page deltas, so a compacted redo
log is not a store — hence heap files, a control file, a redo pointer,
a second recovery source and a separate process
- 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 all of the above from the design
- what IS worth porting is the ordering discipline: publish the new
"recovery starts here" atomically and LAST, so a crash falls back. PG
needs a start-of-checkpoint redo pointer plus an end-of-checkpoint
control file update; we get the same property from one rename, because
we can swap the whole data set atomically and PG cannot
Forks settled:
- no snapshot format — the compacted log is the snapshot, existing grammar,
so no new encoder or decoder and the dump reuses wo_wal_append_insert
- one source, not two
- volume-only trigger, as a ratio against the LAST compaction's measured
output (the denominator is known exactly; estimating the live set would
mean estimating Text) with an absolute floor. NO TIMER — PG's exists to
bound loss from unflushed buffers and we have none; an idle log does not
grow. Copying the mechanism without the reason was the trap
- stop-the-world, with the pause measured against a stated budget rather
than assumed acceptable; alternatives are bought against a number
- compaction may run ONLY where nothing is staged (right after a barrier),
or a staged record lands in a file about to be replaced. Normative
Recorded before it can be found late: compaction invalidates every WAL
offset iteration 2's `resident: keys` stores, so the compactor rebuilds the
offset map as it writes. Nothing breaks today because that storage half is
unimplemented — it would break later, looking like corruption.
Also corrected exploration/postgresql/buffer-and-checkpoint.md, which was
wrong on two counts: PG does NOT update its control file by rename (in-place
full-block write + CRC32C), and its checkpoint sketch assumes writeonce has
segment files, which it does not and deliberately will not.
Grounding measured on master: seed 20000 leaves a 986614-byte log; 20000
updates take it to 2590262 bytes with the SAME live rows, and boot+verify on
that store is 155ms.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
databasev2 4 part A, task 6. Mostly documentation, plus one real fix the
full battery caught.
THE FIX. The drain held EVERY DB reply until the barrier — including
reads, which stage nothing and have no stake in durability. That parked
readers behind an fsync for no reason: durable.sN.mixread.p99 rose from
~1043us to 4057us. Only a statement that actually staged a record now has
its reply held. Caught by the gate, not by review.
THE TRADE, recorded rather than smoothed over. What remains is inherent: a
barrier blocks the owner shard LONGER (more records per fsync) though LESS
OFTEN, so anything queued behind one waits. Three full runs of the same
build gave durable.sN.mixread.p99 of 1043 / 2318 / 4147us and wmix.p99 of
8758 / 20000us — a 2-4x spread with the box near idle. So part A buys ~3x
write throughput at the cost of a longer, noisier tail on the owner shard,
and that is the strongest argument for part B (submit and keep serving).
- durable.sN.*.p99us tolerance widened to 100% WITH the reason in the
code: a 2-4x-variable tail gated at 50% gates the disk, not the engine.
The floor is the real guard and is not slack — mixread's (4172us) came
within 25us of tripping on the worst run. Baseline refreshed; a fresh
full run then passed 106 checks 0 failures
EXIT STATUS MOVED 3 -> 74 (sysexits EX_IOERR). 3 and 4 are already used by
SAMPLES for their own meanings — db-bench's own `verify` exits 3 on a
checksum mismatch, and it is the gate that exercises durability, so a
durability abort exiting 3 would have been indistinguishable from the
mismatch it should help diagnose. The low range belongs to programs.
Docs:
- story: progress, the payoff measured two ways, the cost side, criteria
split met/outstanding, and a "part B — its premise changed" section:
it was justified by "close the 66x gap", but that gap is two problems
and only the concurrent one was a batching problem
- board: standup entry in the six-question shape; both databasev2 4 rows
rewritten. They had said "close the 66x gap" — recorded as MIS-STATED
rather than quietly renumbered
- 00-wob-format.md and 04-db-binding.md: the normative failure contract
("a failed WAL commit traps WO_T_IO after un-applying the row") was
false; corrected, along with the tick-scoped group commit that never
happened
- database/src/CODE-LOGIC.md: where the barrier runs and why there, why
replies are held, why the inline path is asymmetric, the one failure
rule, and how to measure it
- db-bench README: the wmix mode, the env knobs, and the tmpfs warning
Battery: wovm-test 36 suites 0 fail, woc-test, oop-e2e 119/0,
db-bench 106/0, linkcheck clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Plan for the approved spec. Code-free per the repo convention
(docs/plan/discarded.md:54); the executor writes the code.
- T1 a failed barrier is detected and fatal — one entry point that names
the operation, errno, WAL path and batch size, then exits. The abort
path itself stays unexercised and the task says so rather than buying
coverage with a fault-injection switch
- T2 the barrier moves to the drain point and replies are held; the
request path stops committing per append. Riskiest task, and its risk
is one place: the crash legs. Plan says STOP if they fail, do not
adjust the test
- T3 the inline path takes the same fatal rule but keeps its own barrier,
with a comment explaining the asymmetry so the next reader does not
"fix" it. Looks like a no-op; without it the two paths disagree, which
is the unevenness the spec exists to remove
- T4 prove batches actually form BEFORE measuring the payoff — otherwise
a win gets attributed to the wrong cause. Also records peak staged
bytes, settling the no-cap decision with a number
- T5 measure, gate, write it down. If the payoff is absent, say so and
stop: part B must not start on an unproven premise
- T6 closeout, including the error catalogue — WO_T_IO leaving the write
path is language-visible and must be written down
Spec corrected while planning: it pointed at durable.s1.seed as the
payoff. Wrong, structurally — worker shards hold no WAL, so a queue only
exists when other shards write, and a serial writer has nothing to batch
with. The real target is durable.sN.mixwrite: 480 ops/s at p99 5888us
against s1's 1023 at p99 664, so adding shards currently makes durable
writing WORSE. That inversion is a better argument for the iteration than
the one the story recorded.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Brainstormed 2026-08-28. The iteration is split: part A batches, part B
(io_uring submission) is deferred until A's measurement says whether the
blocking boundary still dominates.
The story's premise needed correcting first:
- it says "replace fsync-per-commit with io_uring group-commit", but the
engine commits per STATEMENT — db.c calls wo_wal_commit right after
every append, all six sites, so each row change is one pwrite + one
fdatasync
- so two independent wins were being carried as one, and only the second
needs io_uring. The staging buffer already holds any number of records;
today it never holds more than one. Part A is mostly deleting calls
- iteration 22's numbers say A is where the payoff is: durable writes
4460 ops/s, mixwrite 1023 ops/s p99 664us, against 1.28M ops/s reads
Forks settled:
- batch boundary is QUEUE-DRAIN, not the tick this story had recorded: a
tick adds latency to a lone writer, taxing an idle system to serve a
busy one. Queue-drain self-tunes and needs no knob
- shard 0 holds each reply envelope instead of sending it, commits once
when the queue empties, then releases all — so a writer is acked after
the barrier carrying ITS record, which today is true only because
every batch has one member
- a failure between "RAM mutated" and "record durable" is a FATAL,
diagnosed abort. This replaces uneven behaviour that already exists:
insert rolls back, update and delete do not and say so in a comment
("RAM ahead of disk"). Batching would have multiplied that
- consequence stated, not slipped in: WO_T_IO leaves the write path
- no batch cap initially; peak staged bytes is measured so the question
is settled by a number
One gap disclosed rather than hidden: forcing a real fdatasync failure
needs mount privileges, so the unit test proves the error is DETECTED and
the abort itself stays covered by inspection.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
- docs/stories/databasev2/, numbered from 1. Six PENDING database iterations
moved from the language track and renumbered, keeping the old id in
`was_language_iteration:` so a search for "iteration 32" still finds it:
32 -> 3 WAL checkpoint, 23 -> 4 io_uring commit, 33 -> 7 single-file store,
27 -> 8 query grammar, 20 -> 9 cross-program, 21 -> 10 keypair auth.
Done work (9, 9b, 22) stays as v1 history; language 18 left whole
- the problem, read off the engine not guessed: rows are malloc'd slabs with
addresses stable forever, NO eviction/spill/paging anywhere in database/src,
the WAL never checkpoints so boot replays all history, and durability is one
process-global WO_DATA so no table can say it matters more than another.
An allocation failure IS a clean catchable WO_T_OOM — but swap thrash
arrives first and carries no error signal at all, which is the real hazard
- four new iterations:
1 measure the ceiling FIRST (curve not cliff; the three exits; kill -9 at
exhaustion) — every later default should follow from a number
2 `@table(mode: ram | durable | cold)` — the grammar ask. Small surface
(Ast.table_cfg gains a key, the parser already rejects unknown args), big
semantics: `durable` defaults so nothing changes silently, and the
compiler refuses a durable row holding a `ref` into a ram table
5 bounded tables + refuse/evict/back-pressure, shedding BEFORE the OS acts
6 cold tiering — mostly forks, incl. whether the language surfaces the
fault cost and whether @unique on cold is refused outright. A paged
B-tree stays rejected: if tiering needs one, reject tiering
- 39 links repointed, link TEXT renumbered to track-local ids; arc gains one
pointer row replacing the six moved; board + board-views cover three tracks
- linkcheck 0 broken / 0 anchors; no code blocks in any story
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
