- new chapter 7, "Storage modes: durable and resident", covering what
master gains with the databasev2 cherry-pick: durable: false for a
RAM-only table, resident: keys for a table that outgrows RAM
- states the parts a reader would otherwise hit as surprises: a
keys-resident table is REFUSED at startup without WO_DATA, an update
appends a delta rather than rewriting the row, and the chain is
bounded at 16 links so a hot row does not degrade reads or replay
- quotes the measured 2.55x smaller resident set, not an estimate
- actors/deps/serving shift to ord 8/9/10; seeding is ord-driven so an
existing WO_DATA keeps its rows and only a fresh boot reseeds
- home card says a table can be RAM-only or outgrow RAM
- two gate legs pin the new chapter; site-accept 23 checks, 0 failures
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 3b503c0db50aa737dd06ef6d17151c654a42c59b)
- status: done in the story frontmatter (was the non-conventional
"complete"; the board's axis uses done/in-progress/pending/hold)
- board row 11: what landed, the WO_WAL_UPDATE correction, the ceiling
removed as unreachable, and the one criterion still weaker than
written (expected value, not a resident: all oracle)
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit de39a88e81e97bf6f8b75e9f15331aae20f7ab29)
- flattened row image is WO_WAL_UPDATE, not WO_WAL_INSERT: the row's
original INSERT is already in a live log, so a second one for the same
id is a duplicate replay refuses as corruption. INSERT is right only
for compaction, which builds a fresh log
- remove WO_CKPT_MAX_GARBAGE: with the absolute term at 64 MiB, garbage
large enough to reach a 256 MiB ceiling has already tripped it, so the
branch was unreachable. Postgres needs both constants because it
thresholds on tuples with its pair at opposite ends; this thresholds
on bytes, where one constant does both jobs
- test_delta_chain_flattens_at_k: chain depth stays <= WO_DELTA_MAX_HOPS
across 2K+2 updates, and a reset is observed
- test_delta_chain_flatten_replays: a flattened chain replays correctly
- test_keys_resident_indexed_across_flatten: a delta on an indexed
column composes with flattening, checked at every step across the
bound and after restart. Found no product defect
- test_should_compact_absolute_and_ceiling: pins the absolute term, the
boundary just under it, and the small-log case the ratio still governs
- test_wal 5700 pass / 0 fail; wovm-test and woc-test green
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit f93b5d9db753305c297e868d977670e6d703684c)
TESTS DELIBERATELY HELD at the developer's instruction — logic only.
The existing suite passes (36 suites, 0 fail) but exercises NEITHER new
behaviour: nothing builds a 16-deep chain, and no checkpoint test uses a
log near 64 MiB. Green here means "did not break what existed".
- tier 1: wo_wal_fold_row_at gains hops_out. The walk already visits
every hop, so the depth is free — this is the design's pd_prune_xid,
a cheap "is work worth doing" hint taken from work already happening
- the update path branches on it: past WO_DELTA_MAX_HOPS (16) it writes
a full-row image instead of a delta, terminating the chain. `r`
already holds the complete post-update row because index maintenance
required folding it, so flattening costs bytes, not an extra read
- wo_wal_append_row_image encodes from a caller-held row, as
WO_WAL_INSERT: a chain's base must replay into a database where
nothing precedes it, so replay/compaction/fold need no change
- tier 2: should_compact gains a TRIGGERING absolute term and a ceiling.
Our `floor` SUPPRESSES on a small log — the opposite of postgres's
vac_base_thresh, which triggers on a small absolute problem the
proportion hides. We had the proportion and the suppressor and
neither real guard
- verified by construction, not test: both update entry points converge
on row_apply_field_keys; db.c captures next_offset BEFORE calling in,
so the re-point is transparent to which record type was written
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 1b808abd5942de81c3a6416714d1302384103040)
- new `residency` leg in db-bench.py driving docs/examples/residency-bench:
two tables identical except the annotation, control cap + binding cap
- ITS OWN PROGRAM, not a db-bench mode: declaring a resident: keys table
is a WHOLE-PROGRAM constraint, so the no-WO_DATA refusal fires for
every mode in the module. Putting those classes in db-bench's shared
types made growth/ceiling/randread — which run without WO_DATA —
refuse to start. Caught by running the leg, not by reading it
- gates the RATIOS, waives the absolutes: ops/sec under a cap is swap
and disk I/O and belongs to the box. Same split randread makes
- rss_ratio 2.55 floor 2.0 tol 10% (structural, like bytes_per_row);
overcap_vs_swap_x 1.53 floor 1.0; in_ram_cost_x 4.23 ceiling 8.0;
all_collapse_x 105.4 floor 2.0
- all_collapse_x exists because the leg's FIRST run silently measured
nothing: at QUICK's 40k rows a 48 MiB cap binds neither mode, so the
"over-cap" half was not over cap. The cap now scales with N and the
leg asserts it binds
- verified the gate bites: rss_ratio 1.4, overcap_vs_swap_x 0.6 and
in_ram_cost_x 12.0 are all rejected
- task 7 closed: both criteria moved to Met with how each was verified
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit a310496664982c372f51b43113465eb8ad9e9fb5)
- hreadall/hreadkeys: the same resident A/B as wread_*, but ~2 KB per
row so a GB of data is reachable in a few hundred thousand inserts
- the insert path is fsync-bound at roughly 2 000 rows/s, so row COUNT
is the expensive axis and row SIZE is nearly free — 20k rows already
produce 38 MB
- NOT RUN: the GB-scale measurement was called off. These modes are
committed working and typechecking so the leg can be run later
without rebuilding it, not because a result exists
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit abc276ac39dd45a1052b6ae24d25aedb9eea3e1e)
- two tables identical except the annotation, 200k rows, 40k reads in
one key order, WAL on ext4 (not /tmp, which is tmpfs here and would
have put the log in RAM), rootless cgroup v2 cap
- WIDE shape, 2.55x smaller resident set: 34.4 MB vs 87.5 MB. That is
the real win and the thing the mode was built for
- under a 48 MB cap (between the two resident sets): keys 19635 ops/s
vs all 12854 — only 1.53x faster than letting the kernel swap
- degradation is far gentler though: all collapses 105x from its own
uncapped throughput, keys 16x
- costs 4.2x read throughput when memory is not tight, and writes are
markedly slower — the keys fill did not finish in 2 min where the
resident fill plus 40k reads did. No design doc had costed writes
- THE UNANTICIPATED FINDING: cgroup limits charge the PAGE CACHE, so
moving rows to a file does not escape a container memory limit. WAL
37 MB + RSS 34 MB cannot both live under a 48 MB cap, so every pread
reaches disk. The premise "the page cache will hold the hot rows"
fails in exactly the deployment this targets
- first attempt used Int-only rows and showed parity; recorded, because
drop_payload frees a field's VALUE and an Int's value is its inline
slot word, so that shape cannot benefit and would have condemned the
feature for the wrong reason
- verdict: keep it, to fit ~2.5x more data in given RAM — not to make
an over-capacity table fast
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 7cba9b1174b0bf581314b3147e25cc49e6f49464)
- fixes a limitation iteration 2 shipped: compaction was supposed to
bound chain length, but wo_wal_should_compact triggers on a whole-log
byte ratio and cannot see one hot row's chain
- tier 1, flatten on update: the update path ALREADY folds the row for
index maintenance and the fold already walks hop by hop, so it reports
depth for free. Past a fixed K it writes a full row instead of a
delta. Read <= K+1 reads, replay O(K^2) per row. No format change, no
per-row RAM, no new trigger
- tier 2: our compaction policy has a proportional term and a
SUPPRESSOR misleadingly called a floor; postgres's floor TRIGGERS on
small absolute garbage. Add that term and a ceiling
- design read from .dev/reference/postgresql, not recalled:
heap_page_prune_opt gates on an O(1) on-page hint then page fullness
against Max(fillfactor, BLCKSZ/10); autovacuum uses base + scale *
reltuples clamped by a max (50, 0.2, 1e8). Neither thresholds on
new-bytes-versus-old-bytes
- K deliberately does NOT scale with table size: postgres scales a
table-level aggregate with proportional harm, ours is per-row with
additive cost, so scaling up would make big databases boot worst
- the story says plainly it should NOT be next: task 7 has still never
measured whether resident: keys beats the kernel's own paging
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit f667cad2cfbe187b5973440ab1af015b2df288f8)
- explains replay, the row chain, how a checkpoint flattens it, and why
replay of a long chain is quadratic
- worked SKU example with the actual record layout and back-pointers,
and a trace of the fold showing first-seen-wins
- states plainly why the checkpoint does not bound the hot-row case:
both triggers are ratios over the whole log and nothing counts
per-row chain length
- records the bounded-memory vs linear-time conflict behind the O(N^2)
replay rather than presenting it as an oversight
- closes with the reviewing lesson, since this shape survived several
rounds: complexity bugs hide in the caller's loop, not in the linear
helper being read
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit e6434403d566b5d72a24e9c0dcad1f25a2c16320)
- "no storage behind it" / "nothing yet stores a table that way" was
false — CRUD, checkpoint survival and updates all landed; replaced
with an accurate summary naming task 6/7 as what remains
- the three checked delete/delete-replay/update criteria sat in
Outstanding despite being done; moved to Met, leaving Outstanding
holding only genuine task 6/7 work
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit b575678ee95fa3125fa4e8145320a9cdca10ba38)
- wo_wal_pend_repoint's failure was silently discarded (db.c); its
own doc claimed replay reconciles a stale map — false, a second
same-drain update chains past the lost one, permanently. Now
fatal, like wo_wal_stage_fatal; comment corrected
- apply_delta dereferenced db->rt->wal unguarded — NULL rt + any
DELTA record was a crash. Now refuses cleanly (-1)
- wo_wal_replay_ex lent its throwaway view only when rt->wal was
unset, so a live wal's non-empty staging buffer could be folded
against during replay. Now installs unconditionally whenever rt
exists, saving/restoring whatever was there
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit fed9fe8b5fe022f0b22a4170c7fd68008700939f)
- wo_row_borrow's keys arm folded at hget()'s DURABLE offset even
when an earlier update in the same drain had only a PENDING
re-point
- idx_remove_row then hashed the pre-first-update value, found no
matching bucket entry (already moved by the earlier update), and
idx_add_row added a second one — N same-drain updates leaked N-1
entries, unbounded, nothing reclaims them but a restart
- now prefers wo_wal_repoint_offset1() over the durable offset, same
as back_off already does, closing it for every borrow
- new test: 5 updates to one row in one drain, assert exactly one
index entry — fails (5) before the fix, passes (1) after
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit d4b12d1908e574419c7af2411e52d02623f7e5b7)
- loader stopped refusing durable:true+resident:keys once UPDATE
landed; nothing replaced it at runtime
- rows for such a table live only in the WAL, so every read failed
with a misleading "no such row" instead of naming the problem
- main.c now refuses at startup, names the class, exit(2)
- residency-accept.sh gains a leg: refuses without WO_DATA, still
runs with it — verified failing before the fix, passing after
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 3ea6d6452f260d45f92045c0f300fa49faa1d810)
- loader.c: delete the INCOMPLETE-update BAIL; durable:false +
resident:keys stays refused (nowhere to read from)
- table.c: root-cause fix for the Text-index gap — a keys-resident
borrow now holds ENGINE values, matching wo_row_ptr's contract
(table.h's "no VM pointer" doctrine), not a VM-decoded row. Fixes
idx_hash/idx_cols_equal/wo_idx_probe AND db.c's GET_FIELD/PROBE
arms with one change; reproduced pre-fix as an ASan
heap-buffer-overflow
- docs/examples/residency: Product is genuinely resident:keys;
residency-accept.sh's refusal leg replaced by proving the program
runs and stock survives a restart (11/0)
- test_wal.c: oracle test drives resident:all and resident:keys
through the same update sequence and asserts identical rows;
Text-indexed-update test catches the representation bug; five
pre-existing tests corrected to the fixed contract (4746/0)
- story, README, status board, CODE-LOGIC.md updated; three known
limitations documented: mid-drain stale reads, O(N^2) replay in
chain length, compaction blind to per-row chain length
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit b87c68f950f01aa5e572fbb86a0f374adc83d813)
- apply_delta: DELTA replay arm — fold pre-delta state via back_off,
overlay the field, remove-then-recreate so indexes stay correct
- apply_record/replay loop: dispatch DELTA to apply_delta, drop its
payload back to the log same as INSERT/UPDATE
- stage_flattened_row: compaction's delta-chain path — fold + re-encode
as one fresh INSERT instead of copying the chain
- wo_wal_compact: peek the row's current record kind, flatten deltas,
keep the byte-for-byte copy for chains already at length zero
- test_wal: three new tests — chain-of-three replay incl. secondary
index, compaction flattens to chain length zero (asserts the record
is a full row, not a delta), and the commit-before-repoint crash
window replays the update without ever re-pointing the map
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 7e4ae70d7beb2a2e396a002184bc06f41253decb)
- table.c: unique shadow-check's candidate lookup now checks
wo_wal_repoint_offset1 before the durable wo_row_offset1, same as
back_off — a candidate updated earlier in the SAME uncommitted drain
was folded from its stale pre-update offset, letting a real @unique
clash through and committing a duplicate
- the offset-only substitution alone was NOT enough (verified): the
candidate must be FOLDED to compare values, and folding a pending
offset via pread saw "no record" (bytes still only in the staging
buffer), so the clash was still missed, just for a different reason
- wal.c: wo_wal_fold_row_at now reads a hop inside the currently-staged
region from `w->buf` (new scan_record_staged, scan_record's framing
over memory) instead of pread; every durable hop, and every existing
caller, is unchanged
- test_wal.c: two updates in one drain where the second collides with
the first's new unique value; must be refused. Verified failing
against the prior commit, and still failing with only the offset
substitution, before the fold fix; passing with both in place
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit c049ab92570cfba4d12a018884a20b25a8916727)
- db.c: guard both WO_B_DB_UPDATE_FIELD arms on keys-resident tables —
wo_wal_append_update read a NULL wo_row_ptr there; a live crash, fixed
- ruling override on Task 3: row_apply_field_keys no longer commits or
moves the id map — stages the delta, does the index swap (RAM apply,
unconditional past the shadow-check; a stage failure past that point
is now fatal, like insert). Commit/re-point move to the caller,
mirroring insert. table.c's WAL commit removal is this ruling, not a
regression
- offset passed back via caller-side wo_wal_next_offset(), insert's
koff pattern
- inline arm commits then re-points; request arm records
wo_wal_pend_repoint (own list/name — a drop and a re-point differ),
flushed by wo_db_flush_drops after the barrier
- back_off checks the pending re-point before the durable offset, else
a second update in one drain skips the first delta; verified failing
this way, passing after
- wal.c: fixed a stale comment — keys-resident updates CAN reach
wo_wal_append_update's caller now, they just never call it
- test_wal.c: 2 tests updated for the new contract; new test drives 2
same-row updates via wo_row_update_field_slot in one uncommitted
"drain", checks the value and delta 2's on-disk back-pointer
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 4d13bcebfe51936ba8c608dd9e791c784e5b8983)
- row_apply_field_keys's shadow-check borrowed candidates via
wo_row_borrow, which shares ONE per-table scratch with the row already
borrowed for the update — every candidate borrow returned NULL, clash
was always false, `@unique` silently accepted duplicates on update
- idx_add_row's own internal check has the identical defect at the same
call site; discarding its result is now actually safe, since the fixed
shadow-check clears uniqueness before it ever runs
- fix: extracted keys_fold_into (fold+decode) out of wo_row_borrow so it
can target a throwaway per-call buffer instead of t->scratch; the
shadow-check probes candidates into that buffer — r is never
released-and-reborrowed (r IS t->scratch; that would overwrite it)
- wo_row_borrow itself is behavior-preserving: same checks, same order,
same messages, just factored
- test_wal.c: new test — genuine @unique index, update collides with an
existing row, asserts refusal (DB_ERR_UNIQUE) and both rows untouched;
verified failing (update wrongly succeeded) pre-fix, passing after
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 409186da51fec0042d8d1e3dcf9f69f704937823)
- table.c: wo_row_update_field/_slot no longer refuse `resident: keys`,
both converge on one new static row_apply_field_keys
- borrows (folds), shadow-checks uniqueness, appends the delta with the
row's current offset as back-pointer, commits, THEN idx_remove_row +
idx_add_row + wo_row_set_offset — failure through commit leaves the
row's offset and index untouched
- nv decoded to a VM value before touching the materialised copy, since
wo_row_release drops every slot through the runtime, not db_val_free
- borrow released on every exit, including every failure arm
- resident: all path (row_apply_field_slot) byte-for-byte unchanged;
wal.c untouched — Tasks 1/2 already expose everything needed
- test_wal.c: plain field update read back, and an indexed scalar
column updated then found via wo_idx_probe by its new value, gone
from its old — both verified failing pre-implementation, passing after
- concern: idx_hash/idx_cols_equal/wo_idx_probe cast Text slots to
db_text* unconditionally; a keys-resident borrow decodes Text to a VM
wo_str* (different layout) — pre-existing, left untouched; tests use
a scalar index to sidestep it
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 89c56a13ed9f4abd082bfcabb46f46bb53d39faa)
- wal.c: wo_wal_fold_row_at now refuses any delta back-pointer that
does not point strictly earlier than the record naming it
(back_off >= cur), instead of capping total hops at off/13+1
- this is the real invariant, not a proxy for it: a step-count bound
lets a forward-pointing back-pointer through in one hop whenever it
happens to land on a genuine record, returning a plausible-but-wrong
row instead of refusing it
- removes the 13-byte-record magic number entirely; no arithmetic
tied to record framing remains in the guard
- wal.h: docblock updated to describe the direction invariant
- test_wal.c: two new tests — self-pointing back-pointer (boundary
case, back_off == cur) and forward-pointing back-pointer to a real
future record for the same row (the actual gap: verified failing
against the old step-count guard, passing after the fix)
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 173dbf28a42d45a8c7f9fe430355f61f70c81935)
- wal.h/wal.c: wo_wal_fold_row_at — THE fold. Walks BACKWARD from an
offset through WO_WAL_DELTA records, remembering the first value
seen per field index (newest wins, since newest is seen first),
stops at the first INSERT/UPDATE, decodes it, overlays resolved
fields. Returns ENGINE-owned values so reads, replay, and
compaction (Tasks 3/5) can all build on the same output.
- Cycle guard: caps the walk at what the log up to the starting
offset could possibly hold (13 = scan_record's own record-size
floor), so a corrupt or malicious back-pointer fails loudly
instead of spinning.
- table.c: wo_row_borrow's keys arm now calls the fold instead of
wo_wal_read_row_at directly, then VM-decodes the result — same
two-stage pattern wo_wal_read_row_at used internally. Per-table
scratch, scratch_busy nested-borrow refusal, and the cid/id
identity check all preserved unchanged.
- resident: all path (wo_row_ptr) untouched.
- test_wal.c: two new tests — deltas on two different fields (changed
fields take the new value, the untouched field keeps its original)
and two deltas on the SAME field (the newer wins, pinning direction
— a reversed fold would pass with the older value instead).
Verified failing pre-implementation (wo_row_borrow returned NULL
since a delta record isn't INSERT/UPDATE) and passing after.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit a60231cde1d49d74743cedbd134d2da11158b70b)
- review finding: field_idx=0 and back_off=0 (fresh WAL, offset 0) meant
a u32/u64 swap of these two values wrote identical zero bytes either
way — undetectable by the prior assertions
- test_delta_record: new dedicated 3-scalar-field class (not shared
KEYS_CLASSES) so field_idx can be a nonzero, fixed-8-byte value without
a Text field's variable-length encoding complicating the fixed body
size assertion
- stage+commit a filler row first so the target row's insert record (the
delta's back-pointer) lands at a nonzero offset, not the WAL's initial 0
- delta now targets field_idx=2 with back_off=base_off, both nonzero and
distinct from each other and from class_id=0
- class_id stays 0: this fixture registers exactly one class, so there is
no other value to give it without an unused second class purely to
shift an index
- verified live: temporarily swapped the field_idx/back_off wput calls in
wal.c, confirmed test_wal now fails (fidx==49 want 2, back==2 want 49),
then reverted — wal.c diff is a no-op, only the test changed
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 20ba0965e15e200641b8b63893a2f238a0279fba)
- enum: add WO_WAL_DELTA = 4, existing 1/2/3 untouched (on-disk logs)
- wal.h: document kind 4's payload shape in the format docblock
- wal.h: declare wo_wal_append_delta(w, db, class_id, id, field_idx,
back_off, value) — back-pointer taken as a parameter, not looked up,
keeping the encoder ignorant of table/map state
- wal.c: implement it, modeled on wo_wal_append_insert's shape —
wput_u8/u32/u64 the header fields, enc_val the one field, stage()
- test_wal.c: new test_delta_record — stages a delta after an insert,
commits, then preads the raw record and asserts kind/class/id/
field_idx/back-pointer/value all round-trip; registered in main()
- nothing reads deltas back yet — decode/apply is a later task
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 9c6f832c0534a59e644c53b7cd2850581da12159)
- the plan placed the fold near wo_wal_read_row_at at "line ~600";
it is at line 794
- every other citation verified: wal.h:46, table.c:452/487,
db.c:88/289, wal.c:738, wal.c:861
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit c6cd48679e21e510696391b340c735aa0ad098f9)
- 6 tasks: the record kind, the fold, updates + indexes, the request
path and group commit, replay + compaction, then lifting the loader
refusal and proving it end to end
- the fold is written ONCE and called from three places; the tests are
arranged to prove each caller separately, and the plan says that
wanting a second fold "just for this caller" means the design failed
- Task 2 includes a same-field ordering test specifically, because a
fold walking the chain backwards the wrong way returns plausible data
and is otherwise invisible
- Task 6 step 1 audits the db.c request arms BEFORE lifting anything —
they were never audited for keys-residency the way the inline path
was, and the last audit of that kind found delete corrupting memory
- self-review found the spec's crash criterion had no task: added a step
that commits a delta, skips the re-point, and replays, which is the
state a crash between barrier and flush leaves behind
- every symbol the plan names verified to exist in database/src
- code-free per house convention; the writing-plans skill wants code
blocks and the project rule overrides it
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit abb8fc9454bdca9d8d703c8cfeb224f89a93d14d)
- updates append a DELTA (class, id, field, value, back-pointer), not a
full row. The workload decides it: a product catalogue changes one
narrow field of a wide row on every order, so a full-row append would
rewrite every field to move one integer on a shop's hottest path
- the back-pointer keeps the id map at one slot per row, which is the
mode's whole premise; a map growing per update would defeat it
- ONE fold function, three callers (read, replay, compaction). Three
implementations of one rule is how they drift, and a fold that differs
between reading and replaying is a database that changes its mind at
boot. Named as the design's principal risk
- indexed columns MAY change: price is exactly what a catalogue indexes,
so forbidding it would be a restriction users meet immediately
- no chain cap, deliberately. Compaction already rewrites live rows, so
every checkpoint resets every chain, and deltas grow the log which
pulls the next checkpoint forward — the workload that lengthens chains
triggers the fold that shortens them
- the risk that accepts: one hot SKU under an otherwise quiet write
rate. Task 7's benchmark must include it
- supersedes the 2026-08-26 spec's one-line full-row Update sketch,
marked in place rather than left as a second design in the tree
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit c9a88b05c4c62a5df13253686c990aaccc3f9cbb)
- the motivating workload was an audit log, which is append-only and so
argues for nothing. A catalogue is the real case: stable ids, and
stock moving on every order while name/price/sku sit still
- Product (durable, resident: all) and Cart (durable: false), with
place_order decrementing stock through a write-through field assign
- run `order` twice and stock goes 10 -> 7 -> 4: a level below the
seeded 10 can only mean an earlier order's UPDATE replayed. That is
the stronger claim — not just that inserts survive, but that a field
change does
- caught by running it three times: my first assertion required
before == 10, which only holds on a fresh seed and failed on the
third run even though the data was correct
- gate gains a leg for the update-replay claim; residency 12/0
- the commented resident: keys block now argues the DESIGN too: only
stock changes per sale, so appending the whole row would rewrite
every field to move one integer on a shop's hottest write path
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit d4104dcc5e3a460459dbf2d24043ce25b113bcdf)
- wo_row_remove's keys arm borrows the row from the log to find its
index entries, and a borrow reads through db->rt->wal. At boot that
pointer is not wired yet: main.c replays first (main.c:226) and
assigns rt.wal afterwards (main.c:268)
- so the borrow found no log, the remove failed, and replay reported a
valid tombstone as CORRUPTION. An UPDATE record would have failed the
same way, since replay applies it as remove-then-recreate
- replay now lends the runtime a read-only view over the fd it already
has open, for the replay's duration only, and restores what was there
- broken by the delete fix in 76b8fd9 — deletes worked in-process but
their tombstones broke the next boot. Unreachable in production only
because the loader still refuses the annotation
- pinned by test_keys_resident_delete_then_replay, verified failing
against the unfixed code (2 failures) and clean with it
- found by asking whether the read-modify-append plan was ready, not by
a gate
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit dc25462461b9f79d70c803f7174adc90fa16c90e)
- docs/examples/residency: one program, two tables filled by the same
loop, differing only in the annotation. Run twice against one WO_DATA
and orders replay while sessions do not
- the example checks its own claim (exits 1 if a durable:false table
survives, or a durable:true one fails to replay) rather than narrating
it in a print
- resident: keys is written out as a commented block with the loader's
exact refusal, so the frontier is visible in the example rather than
only in a story. It documents WHERE the refusal happens: woc compiles
it and emits a .wob; wovm exits 2, because the annotation is a
load-time property
- residency-accept gains two legs: the example runs and its restart
claim holds, and the refusal message the README quotes is checked so
doc and code cannot drift apart
- the gate writes the example's output to /tmp/residency.log,
banner-separated, for tail -F
- README commands verified verbatim; they needed mkdir -p because wovm
will not create WO_DATA
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit c9c7e03e62c3918cb65ef7994d1e33a0c5337b71)
- wo_row_remove read the id map's value as a slot, but on a keys table
that value is a LOG OFFSET (hput(t, id, wal_off + 1)). slot_row does
no bounds check, so a delete indexed t->slabs[] with a byte offset and
then called db_val_free on whatever it landed on — arbitrary frees,
not a wrong answer
- keys tables now take their own arm: no slab slot, no bitmap bit, no
free-list entry to return. The index hook needs the row's values, so
the row is borrowed from the log for exactly that long
- wo_row_ptr carried the same trap and is public. It cannot refuse keys
tables outright (insert legitimately calls it while the map still
holds a slot), so it now detects the offset case — index past the
slabs, or bitmap bit clear — and returns NULL. Callers all handle NULL
- test_keys_resident_delete pins it; it SEGVs against the old code,
verified by reverting the fix rather than assumed
- found while auditing every hget() reader before narrowing the loader
refusal to allow benchmarking. The refusal was justified in the docs
by "updates are unimplemented" while actually standing in front of
this too: a guard whose stated reason is narrower than its real one
gets removed by someone who believes the stated reason
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 76b8fd944af9ed062467bdf9ab93c2e96dd198cf)
- chain is a cross-track field (1-6); only databasev2 3 and 4 carry it,
and iteration 2 — critical path, in-progress — carries none, so the
board's chain query cannot see it
- 00-story.md's ASCII graph draws 3 before 4; the chain field and the
history both say 4 first (4 part A 08-28, 3 08-29). Graph is wrong
- graph also contradicts its own prose on edge direction, and still
draws the 2-5-6 path the 2026-08-27 amendment retired
- iteration 3's hazard section is stale: it says nothing fails "because
iteration 2's storage half is unimplemented", which 5c/5d ended
- and it was incomplete: it named offsets going stale, but compaction
walked the bitmap, which a keys row has no bit in, so those rows
would have been dropped from the new log outright — data loss, not a
bad pointer, and offset-rebuilding would not have caught it
- coupling is now bidirectional: wal.c compaction calls iteration 2's
wo_row_next_id / wo_row_offset1 / wo_row_set_offset
- nothing in the track is blocked on anything else in the track
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 2ecaf0c95f40c6d6c8cc2fde687c82d63cda1f7e)
- loader's resident:keys refusal said "rows are still fully resident"
and "until tasks 5c/5d land". Both false since f606fc9. Corrected to
name the real blocker: UPDATE needs read-modify-append
- databasev2 00-story: the sequence graph drew 2->3->4, which reads as
3 needing 2 and 4 needing 3. Both backwards, and it still drew the
2->5->6 path the 2026-08-27 amendment retired. Redrawn stating only
real dependencies, with 4 and 3 shown as composing rather than
ordered, and the execution order that actually happened
- databasev2 03: the hazard and its Outstanding entry both claimed
nothing fails "because iteration 2's storage half is unimplemented".
Marked discharged, and recorded that the hazard named only half the
danger — the bitmap walk would have dropped keys rows outright
- databasev2 06: pending -> hold (largely superseded, revisit only on
a measurement); dated its 5c/5d references
- porch 01: rewritten to the settled shape. readiness ready, status
in-progress, phases B and C marked superseded with why
- porch 01 claimed time.after "is still a reserved builtin id". False —
builtin 90, implemented. That claim is what made the iteration look
cheaper than it is
- porch README gains honest ledger rows for both features (partial,
being rebuilt), not shipped
- skill-catalog README pointed at a story path that moved tracks;
linkcheck now 0 broken
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit b3d8c403e1d19ac27ec966de85cb293e0765795c)
- wo_row_read and the @unique shadow probe go through borrow/release;
release runs before every exit, including wo_row_read's early return
- updates on a keys table refused explicitly in wo_row_update_field and
the slot variant: no slab slot to mutate, and writing the borrow's
scratch would discard the write silently. Needs read-modify-append
- compaction walked the bitmap, which a keys row has no bit in — every
such row would have been dropped from the new log. Now walks
wo_row_next_id and re-points each row to where it lands
- moves records byte-for-byte (copy_record) rather than decoding: a
borrowed row holds VM values, enc_val expects engine values, and ASan
caught that mismatch as a 4294967292-byte memcpy
- wo_row_set_offset updates a value in place and never rehashes, so a
wo_row_next_id cursor stays valid while compaction re-points
- a compaction that fails after moving rows is fatal: the map would name
an unlinked temp file, and the intact log replays correctly
- test_keys_resident_survives_compaction pins both failure modes; rows
rewrite in hash order so offsets really move
- loader still refuses resident: keys — updates are not implemented
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit f606fc9b76d983cac2b348f03f7d4f01433cd905)
databasev2 2, task 5d. Every reader in db.c now works for both backings.
The measured problem: a keys-resident table's bitmap is EMPTY by construction
(its payloads live in the log), so all four bitmap walks would have silently
returned no rows — a query over such a table would find nothing, with no error.
- wo_row_next_id: one iterator, two backings. Keys tables walk the id map;
resident tables keep walking the BITMAP deliberately, because the id map
holds the same set in hash order and switching would reorder the results of
every unordered query in the repo. No behaviour change where none was needed
- the three id-collecting scans move onto it. They only ever collected ids
(the 9b cursor-stability rule materialises the list up front), so they needed
no row access at all — which is why this was far smaller than the plan feared
- the two filtered scans borrow, compare, and RELEASE BEFORE any exit. The
scratch is per-table, so a borrow leaked past a `return` or `break` would
make the next borrow on that table fail as a nested one. That is a real
hazard, not a hypothetical: the request-path GET_FIELD borrowed and then
`break`ed without releasing until this commit
- point reads decode or clone BEFORE releasing, because a keys-resident row's
slots point into the scratch that release frees
Verified: just wovm-test — 36 suites 0 fail.
Still to do in 5d: table.c's unique shadow and its three remaining wo_row_ptr
sites, wal.c's append encode, and compaction's own walk — which is where the
recorded `resident: keys` offset obligation has to be honoured. The loader
refusal stays until all of it lands.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 0c97fa48d3e31ea9eea3287d9c23ed29f0260488)
databasev2 2, task 5c. The write and boot halves. Still not exposed: the
loader refuses `resident: keys` until 5d rewires the readers.
GROUP COMMIT FORCED THE DESIGN. A keys-resident payload can only be dropped
once its record is durable, but databasev2 4 deferred the barrier to the drain
— so at append time the bytes are still in the staging buffer and the recorded
offset would pread ZEROS. Dropping at append would have produced rows that
read as garbage, intermittently, only under multi-shard load.
So the drop is recorded, not performed:
- wo_wal gains a pending-drop list, the same shape as the drain's held replies
and for the same reason
- both write paths take the offset BEFORE the append (wo_wal_next_offset) and
record it; the inline path flushes right after its own commit, the request
path's flush runs in the drain immediately after the barrier
- if the process dies before the barrier the list dies with it, which is
correct: nothing was dropped and nothing was lost
- an out-of-memory pend is ignored on purpose — the row simply stays resident,
which is safe
Boot: replay now leaves a keys-resident table pointing at the LOG. Each record
is applied normally, so indexes and uniqueness are built exactly as for any
other table, and the payload is then dropped with THAT record's offset. For an
update the later record wins, because each apply overwrites the map in order —
the rule replay already follows.
Tests: the round trip (insert, commit, drop, read back with Text intact) and
now BOOT — a fresh wo_db replays the store and every row materialises from the
log, count intact, nothing in a slab.
Verified: just wovm-test — 36 suites 0 fail, test_wal 4301 pass, cli_smoke OK.
REMAINING (5d), and precise: every reader still goes through wo_row_ptr, which
for a keys table would index a freed slot. The scans in db.c walk the BITMAP,
and a keys table's bitmap is empty by construction — so a query over one would
today return no rows at all. That, FK restrict, and the @unique shadow are 5d,
and the loader refusal stays until they land.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 08abd09bf88918f2582e74713dc7903beb8aaeb8)
databasev2 2, task 5c step 2. The storage half the accessor was left waiting
for. Not yet wired into insert (that and 5d remain), and the loader still
refuses `resident: keys`, so nothing is exposed to a program yet.
- wo_db gains an `rt` back-pointer, set in main.c beside VM.rt.db. wo_rt
already carries `db` and `wal` as opaque handles, so this closes the loop
and a borrow can reach the log WITHOUT threading a wal pointer through
eleven call sites — which is the whole reason 5c is one accessor
- wo_row_drop_payload: the operation the plan recorded as MISSING. Frees the
slot and its engine-owned values, then re-points the id map at the record's
log offset (off + 1, reusing the same 0-is-empty trick as slot + 1). It
deliberately does NOT touch the secondary indexes (they store row ids, so
they stay correct), does NOT decrement count (the row is still live, only
its backing moved), and does NOT remove the id (that is how it is found)
- wo_row_borrow materialises for a keys table: reads the offset from the id
map, calls 5b's wo_wal_read_row_at into the per-table scratch, and checks
the record actually holds the expected class and id — a compaction that
moved records without rebuilding the map lands exactly there, which is the
obligation recorded at wo_wal_compact
- fully-resident tables keep today's path and pay one predicate
A REAL BUG, exposed the first time the path was used: wo_row_release freed the
materialised values with the ENGINE's allocator. They are VM values —
wo_wal_read_row_at is the out-gate and always copies — so ASan reported a
bad-free immediately. It now drops them through the runtime. That stub was
written in 5c step 1 for a path that did not exist yet.
Recorded while implementing: wo_wal_next_offset's contract says to trust an
offset "only after the matching commit returns 0". Group commit (databasev2 4)
defers that barrier to the drain, so db.c can no longer check inline — but part
A also made a failed commit FATAL, so no execution can record an offset whose
record never became durable. Same guarantee, different mechanism.
Test: a heap-valued row is inserted, committed, has its payload dropped, and is
read back out of the log with its Text intact; count is unchanged (still live);
and a second borrow succeeds, which fails if release did not clear the scratch.
Verified: just wovm-test — 36 suites 0 fail, test_wal 4273 pass, cli_smoke OK.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
(cherry picked from commit 125bd09218d616b2a16b140de770d3f38b45f0ac)
The branch was 17 ahead / 25 behind with 11 conflicting files, and drifting
further: db.c had been rewritten twice on master since (group commit, then
compaction). Resolved rather than rebased so both histories stay legible.
Conflicts, and how each was settled:
- db.c: BOTH semantics kept. Master's fatal path and compaction check now sit
behind the branch's `table_is_durable` predicate, in all three inline arms —
a volatile table reaches neither the barrier nor the compaction check
- db-bench sample: every mode from both sides (growth, growth-verify, randread,
replayseed, wmix) and ONE `boot` mode, which both sides had added
independently
- db-bench.py: all six legs kept. Both sides had also grown the same
WAL-size helper under different names; collapsed into one
- perf-targets: the branch's §5 (RAM ceiling) then master's §6/§7 — master's
numbering had already assumed a §5 it did not have
- story frontmatter: master's `status` (the landing truth) plus the branch's
`readiness` axis. 03 would have read `done` + `refine`, which is a
contradiction — it was brainstormed and landed on master, so `ready`
- board: both standup blocks newest-first; master's chain rows (a superset);
the branch's databasev2 1-2 rows with master's 3-4. Fixed a stray `|` in
master's row 3
- baseline: master's, then REGENERATED from a full campaign — 143 metrics,
132 checks, 0 failures with both sides' legs present
TWO HALF-EXPOSED FEATURES FIXED, because the merge rule is that master gets
no feature that is honoured in name only:
- `resident: keys` PARSED, set a .wob flag, and did nothing: rows stayed fully
resident. A developer could declare a 120 GB table keys-resident, watch it
compile, and be OOM-killed. The loader now REFUSES it with a message naming
what to write instead, until tasks 5c/5d land. The compiler still parses it
and its AST golden still passes, so the grammar work stays tested
- `durable: false` was honoured ONLY on the inline path. wo_db_exec_req had no
guard at all, so a volatile table written from an actor on a worker shard
would still be logged — precisely porch's session-table case, and precisely
what iteration 2 exists to provide. All three request-path arms now carry the
same predicate. Found by reading the merged code, not by a test: the obvious
probe runs main() on the primary and therefore only exercises the inline path
Verified on the merged tree: wovm-test 0, woc-test 0, oop-e2e 122/0,
residency-accept 8/0, db-bench 132/0, linkcheck clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
databasev2 3, task 6. Documentation, plus three gate-tolerance
corrections that are justified rather than silent.
- 04-db-binding.md: the NORMATIVE rule — compaction may run only where
nothing is staged (a correctness requirement, not scheduling), recovery
is unchanged, and a failed compaction is a missed optimisation rather
than a durability event
- database/src/CODE-LOGIC.md: why one file and not snapshot-plus-tail
(Postgres CANNOT compact — page deltas; ours are full row images, so a
compacted log IS a store), why rename is the whole crash-safety story,
why the dump flushes but does NOT fsync when it does, why the
replacement is preallocated, and where the trigger is checked
- README: the checkpoint knobs, the extended walstats line, the boot mode
- story -> status: done, with criteria split met/outstanding
- board: standup entry in the six-question shape, both rows rewritten
THE OBLIGATION IS AT THE COMPACTOR, not only in a spec: compaction moves
every record, so it invalidates every WAL offset iteration 2's
`resident: keys` stores, and the loop that knows each record's new
position must rebuild that map. Nothing fails today because that storage
half is unimplemented — it would fail later, looking like corruption.
Board claim corrected before it shipped: I wrote that the concurrency
chain is "complete". It is not — chain 5 stays in-progress because
databasev2 4's part B was never done and its premise was invalidated by
part A. Every link has landed its PLANNED work; that is a different
statement.
Gate tolerances, each with the measurement that justifies it:
- ckpt.pause_us_max is no longer gated relatively. The raw pause scales
with the live set and this workload's live set is not fixed (wmix's
hist_dump inserts a row per latency bucket), so gating it gates the
box. Added ckpt.pause_us_per_mb — the engine's own rate, gated for
real, and the metric that would have caught the 8x dump regression —
with the absolute 50ms budget still guarding the raw pause
- ram.*.msgrate 15% -> 70%. PRE-EXISTING, and measured: 10.7M-17.9M
msgs/sec across ten full runs, several predating this work — a 1.67x
spread against a 15% gate
- durable.sN.*.p99us 100% -> 300%, with more evidence than the first
widening: mixread 1043/2318/4147us, mixwrite 1623/4446us on the same
build. Floors stay the real guard and are not slack
Battery: wovm-test 36 suites 0 fail, woc-test, oop-e2e 119/0,
db-bench 117 checks 0 failures, linkcheck clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
databasev2 3, task 5.
Full campaign, same workload twice, differing only in whether
checkpointing may fire:
- WAL used 1962358 -> 907094 bytes (2.16x reclaimed)
- boot 114 -> 64 ms (1.78x), median of 3
- stop-the-world pause max 2651us against a STATED 50ms budget
The budget is asserted, not assumed: 50ms is a stall a serving process
can absorb without a client seeing a timeout, and the leg fails if it is
exceeded. The pause is O(live rows) — at ~181 MB/s a 1GB live set implies
~5.5s, which is the number an incremental design must be bought against.
The spec deliberately did not buy it in advance.
FOUND BY MEASURING: the dump was 8x slower than it needed to be. It
flushed through wo_wal_commit, which fdatasyncs, so it paid one barrier
per 256 records. Intermediate durability there is worthless — the temp is
not authoritative until the rename and is fsynced once immediately before
it. With a single final barrier:
- ~107KB live: 23948us -> 2903us
- ~500KB live: 36361us -> 7526us
- ~1.98MB live: 107649us -> 13212us
- marginal ~22 MB/s -> ~181 MB/s, sync-bound to bandwidth-bound
Correctness re-proven after that change: wovm-test 36 suites 0 fail,
test_wal 760 pass including the 40-round kill-during-compaction battery.
Two measurement defects of my own, fixed rather than reported:
- boot measured through the driver's run() helper reported 251ms both
with and without checkpointing — run() samples RSS on a 250ms poll, so
every timing floors at the quantum. Measured directly instead, median
of 3
- ckpt.reclaim_x was recorded as lower-is-better by the default detector,
which would have PASSED "reclaimed nothing" and FAILED an improvement:
the feature's central claim, gated backwards. Now higher-is-better,
gated at 15% while the wall-clock metrics stay wide — waiving them all
would have left the leg ungated, part A's task 4 mistake
- sample gains a `boot` mode that does nothing, so boot time is boot time
- walstats now reports compactions, pause max/total and compacted bytes
- baseline refreshed from the FULL campaign (N=20000, crash_reps=3), and
a fresh full run passes 116 checks 0 failures
- gate bites: reclaim_x doctored to 1.0 -> FAIL on exactly that metric
One flake seen and checked, not papered over: durable.sN.query.ops_sec
failed once at 53% below baseline. It is a read-only metric that touches
no WAL code, and a re-run passed 116/0 with the box at load 1.85.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
databasev2 3, task 4. The plan called this the riskiest task because a race
can pass by luck, so it is argued with mutants rather than green runs.
The battery: a forked child inserts, acks, deletes the oldest so HISTORY
grows while the live set stays ~17, and compacts every 24 iterations. The
parent SIGKILLs at varied instants so kills land before, inside and after
rewrites, then replays and checks the ACKED LIVE SET. The existing
battery's "records >= acks" oracle cannot be reused: collapsing history is
exactly what compaction is for.
A REAL DEFECT IN MY FIRST VERSION, found by the failures and fixed in the
TEST, not by weakening it:
- the child acked deletes AFTER committing them, so a kill in between left
the row legitimately gone on disk while the last ack still said
"inserted" — the parent then demanded a row the engine was right to
remove. Symptom was an acked insert missing near the end of the stream,
~1 run in 3
- deletes now announce INTENT BEFORE committing, so such a row's fate is
simply UNKNOWN to the parent, which is the honest thing to assert. Every
acked insert never marked for deletion must still be present with its
acked value
- the stale-temp assertion was also wrong: it checked for absence after
wo_wal_replay, which never opens the WAL. The guarantee is "removed AT
OPEN", so the test now opens and then asserts. A temp surviving a kill
is expected debris, not a defect
Proven to have teeth, which matters because assertions were softened:
- against the design's rejected alternative (in-place rewrite instead of
the atomic rename) it fails EVERY run, reporting log_records=0 — the
kill landed mid-copy and destroyed the log. That is the corruption
rename exists to prevent
- on correct code: 10 consecutive runs x 40 rounds clean, plus the suite
Also carried the log's PREALLOCATION to the replacement. The WAL is
preallocated so appends never extend the file, which is what lets
fdatasync alone be the ack barrier; a replacement opened with prealloc 0
silently changes that property and the zero-padded tail the scan relies
on. Stated honestly: this is hygiene making the replacement equivalent to
what open() would have produced — I could NOT prove it was the cause of
the observed loss, and the ack race above explains it.
Verified: just wovm-test — 36 suites 0 fail, test_wal 760 pass, cli_smoke OK.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
databasev2 3, task 3.
- wo_wal_should_compact is a PURE decision (used bytes, last compaction's
measured output, floor, ratio) so it is testable without a store —
which is the only way a policy like this gets tested at all. Denominator
is the last compaction's real output, not an estimate of the live set:
estimating would mean estimating Text
- 8 boundary assertions incl. "exactly 3x is not MORE than 3x" and a zero
ratio disabling the policy rather than dividing by nothing
- MUTATION-TESTED instead of observing RED: implementation and test were
written together, so removing the floor check was verified to fail
exactly the two floor assertions. Equivalent evidence, stated plainly
- WO_CHECKPOINT_BYTES / WO_CHECKPOINT_RATIO at boot beside WO_MAILBOX.
The knobs are what make the policy testable — a gate sets a tiny floor
and forces compaction in a few writes instead of megabytes
- NO timer, per the spec: Postgres' CheckPointTimeout bounds loss from
unflushed buffers; our records are durable at commit and an idle log
does not grow
- the ordering rule is now asserted, not trusted: a test stages a record,
requests compaction, and requires REFUSAL with the log untouched and
the staged record still committable afterwards
FOUND AND FIXED a gap in my own wiring. The plan said to call the check
"after the drain's barrier", and I did — but a statement running ON the
owner shard never enters that drain, so WO_SHARDS=1 never compacted and
its log grew forever: measured 536086 bytes where the multi-shard run
held 446024. Now checked after the inline path's commit too (db.c
maybe_compact), where the buffer is equally empty. WO_SHARDS=1 went
536086 -> 260657 bytes. For a checkpoint this mattered more than part A's
equivalent gap: an unbounded log is an operational failure, not just lost
throughput.
Also corrected a measurement of my own: multi-shard logs looked unbounded
(448KB -> 1013KB -> 1647KB across 8k/24k/48k updates). They are not.
Instrumentation showed compaction ran 25 times with zero failures, each
writing MORE than the last, because the live set genuinely grows — wmix's
hist_dump and done-markers are themselves durable inserts. Final log
1631040 against a last compaction of 866432 is a ratio of 1.88, just under
the 2x threshold: the policy holding exactly.
Replies are released BEFORE compaction runs, deliberately: their records
are already durable, and holding them across a stop-the-world rewrite
would add its full duration to their latency for nothing.
Verified: wovm-test 36 suites 0 fail, test_wal 360 pass; db-bench-quick
crash.s1/crash.sN and both restart legs green, and part A still batches
(sN mean 4.16, peak 30).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
databasev2 3, task 2.
- wo_wal_open removes `<log>.compact` before reading anything. The only
way one exists is a crash before the rename, which means its records
were never authoritative
- deleted rather than ignored, deliberately: a file full of well-formed
records sitting beside the log is exactly what a future reader mistakes
for data
Test uses PLAUSIBLE content, not garbage — a byte copy of a real log —
because garbage would be rejected by the CRC anyway and would prove
nothing. It asserts the temp is present before the open, gone after, and
that the live log still replays to exactly what it said.
RED was an assertion failure (`access(tmp, F_OK) != 0` unmet), not a
compile error, so the test was proven to exercise the behaviour before the
behaviour existed.
Verified: just wovm-test — 36 suites 0 fail, test_wal 340 pass, cli_smoke OK.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
databasev2 3, task 1.
- walks each class's live rows via the bitmap-over-slabs pattern db.c
already uses in three places, appending one INSERT per live row through
the EXISTING append path. No second encoder, no new format, and ids are
preserved exactly because wo_wal_append_insert takes the id and reads
the row from the store
- FLUSHES EVERY 256 RECORDS rather than staging the whole store: stage()
grows the staging buffer by doubling and never shrinks it, so a
one-buffer dump would hold the entire store in RAM on top of the store
— the unbounded growth databasev2 1 identified as how this engine dies
- the switch, in order: fsync the temp file, rename over the live path,
fsync the PARENT DIRECTORY (rename's atomicity is in-kernel; the
directory entry is not durable until the parent is synced — Postgres
does the same for the same reason), then reopen the descriptor, because
the old one refers to an unlinked inode
- REFUSES when anything is staged: those records would land in a file
about to be replaced. The caller-side guard is task 3; this is the
backstop
- a failure leaves the ORIGINAL log intact and usable and returns -1. A
failed checkpoint is a missed optimisation, not a durability event, so
it deliberately does NOT take databasev2 4's fatal path
- records the bytes written, so task 3's trigger can compare against a
measured denominator instead of estimating the live set (which would
mean estimating Text)
Recovery is untouched — the result is an ordinary log in the ordinary
grammar, replayed from byte 0. Crash safety comes from rename, not from
code of ours.
Test asserts BOTH halves, on purpose:
- the log shrinks: 43 records (3 inserts + 40 updates of the SAME row, so
history grows while the live set does not) -> 3 records, fewer bytes
- AND a fresh replay reproduces the store: every id present, and row 0
carries the 40th update's value rather than its original. "It got
shorter" is also true of a truncating bug, so the replay comparison is
what actually proves it
- and the WAL stays usable after the swap: a further append lands after
the compacted records, giving 4 on the next check
Verified: just wovm-test — 36 suites 0 fail, test_wal 315 pass (was 165),
cli_smoke OK.
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 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>
databasev2 4 part A, task 5.
Controlled before/after — same machine, same workload (wmix 4000 32),
same build except db.c and vm.c, two runs each interleaved:
- per-statement barrier: 2213 / 2177 ops/sec, p50 7183 / 7251us
- group commit: 6216 / 6525 ops/sec, p50 3458 / 3444us
- ~2.9x throughput, ~2.1x lower p50
The full campaign confirms it a second way: s1 takes the inline path and
commits per statement BY DESIGN, so within one build the shard configs are
batching-off vs batching-on — 1467 -> 5117 ops/sec, mean batch 1.0 -> 5.43,
peak 1 -> 57. 3.5x, agreeing with the 2.9x above.
Recorded honestly:
- the BEFORE p99 is at the histogram ceiling (hist_add clamps at 20000us
and both runs pinned there), so the true figure is >=20ms and unknown.
The improvement is AT LEAST 2.3x; the old p99 was off the instrument
- durable.sN.mixwrite went 480 -> 492 ops/sec, i.e. UNCHANGED. That was
the spec's original payoff metric and correcting it was part of the
brainstorm: mix performs 20 writes at C=4, mean batch 1.01. A workload
that never has two writes in flight cannot be helped by batching them
- seed is likewise unchanged: a serial writer has nothing to batch with
- so the payoff is real but CONDITIONAL — it appears where concurrent
durable writes fan into the owner shard, and nowhere else
Two traps recorded in perf-targets §6:
- do not benchmark durability on /tmp: it is tmpfs here, where fdatasync
is free. The same run reported 195000 ops/sec at p50 1us there against
2200 at p50 7200us on ext4 — no barrier to amortise, so the measurement
measures nothing. db-bench keeps its stores under bench/ for this reason
- the record count is not the update count: 7755 records for 4000 updates,
because hist_dump and the done-marker are themselves durable inserts
- FIXED a regression I introduced in T4: master's committed baseline is
FULL mode (N=20000, crash_reps=3, msg_n=200000) and I had overwritten it
with quick-mode values. Regenerated from a full campaign; the full run
now passes 106 checks 0 failures against it
- gate still bites: sN wmix ops_sec -70% -> FAIL on exactly that metric
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
databasev2 4 part A, task 4. Scope extended with developer approval: the
plan authorised touching the sample only for observability, but no
existing leg has enough concurrent durable writes to exercise group
commit at all, so the payoff was unevaluable either way.
The finding that forced it:
- `mix` writes on one op in ten with C=4 (all_mode calls mix_mode(n/10,
4); Mixer writes on i % 10 == 9), so the quick run performs 20 writes
total. Measured mean batch 1.01 over 3112 barriers, peak 3
- that is a property of the WORKLOAD, not the mechanism: peak 3 of a
possible 4 shows batches form whenever writes actually coincide
- `wmix N C` added: every op a durable write, C at once. Updates rather
than inserts, so it is comparable to mixwrite and the row count stays
flat. Histogram kind 2 — a replayed store still holds the seeding run's
kind-0/1 Hist rows and merging those would report someone else's
latencies
- WO_WAL_STATS=1 prints one line at exit: batches, records, peak_batch,
peak_staged. Opt-in, because it would otherwise pollute every durable
program's output. Counters live in wo_wal; no builtin, the numbers are
diagnostic and not part of the language
Measured, and it scales with concurrency exactly as designed:
- C = 4 / 16 / 64 -> mean batch 1.13 / 1.76 / 5.35, peak 3 / 10 / 39
- the gate's own legs: durable.s1 5412 records over 5412 barriers (mean
1.0, peak 1 — the inline path, one barrier per statement BY DESIGN),
durable.sN 7757 over 2296 (mean 3.38, peak 28) at 2x the throughput
- peak staged 1372 B settles the no-cap decision with a number: the batch
is tiny, so the upstream mailbox bound is sufficient
- mean_batch/peak_batch are higher-is-better (the default detector would
have called bigger batches worse)
- only the batch SHAPE metrics are waived to 100%; wmix throughput and
latency keep real tolerances (15% s1, 50% sN) — a blanket waiver would
have left the entire new leg ungated
- the live assertion `mean > 1.0` on the sN leg is what catches inertness
- gate bites: sN wmix ops_sec -60% -> FAIL on exactly that metric, 1 of 86
Verified: db-bench-quick 89 checks 0 failures; baseline refreshed (86
metrics).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
databasev2 4 part A, task 3. Looks like a no-op; it is not — without it
the two write paths would disagree about what a failure means, which is
the unevenness the spec exists to remove.
- inline path (a statement already on shard 0) keeps its own barrier,
batch size 1. It cannot hold a reply: it returns into its OWN fiber
rather than unparking a requester, so batching it would need that fiber
parked on the barrier — part B's machinery, deliberately out of part A
- the comment says so, and says why not to "fix" it, because the next
reader will otherwise see an inconsistency and delete the commit
- the ordering assumption is written down: committing here is safe only
because the drain commits unconditionally whenever anything is staged,
so the buffer is empty when this runs. If that stops holding, this
commit would make another statement's record durable early and ack it
to the wrong writer
- staging and commit failures are fatal here too. The update arm's old
comment admitted what it did — "RAM ahead of disk: trap, do not ack" —
and that is now gone
WO_T_IO no longer appears anywhere in db.c: the write path cannot be
caught. Language-visible, and task 6 records it in the error catalogue.
Verified:
- just wovm-test: 36 suites 0 fail, cli_smoke OK
- WO_SHARDS=1 db-bench-quick: 85 checks 0 failures, crash.s1.0 800 acked
rows present after kill -9 — the configuration that takes this path
exclusively
- default shards: 85 checks 0 failures
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
databasev2 4 part A, task 2. The core change, and mostly deletion.
- the REQUEST path (wo_db_exec_req) no longer commits after each append.
Applying to RAM and staging stay exactly where they were
- wo_vm_adopt holds each DB reply envelope in a local FIFO instead of
pushing it as the statement finishes. Pushing there would unpark the
requester before its record is durable — the ack contract this
iteration exists to make literally true rather than true by accident
of every batch having one member
- at the end of the drain: ONE wo_wal_commit_fatal for everything staged,
then every held reply. Locals rather than per-shard state: nothing
needs to outlive the batch it describes
- "did this statement stage anything" is asked of the buffer, not guessed
from the opcode, and that count is what the failure diagnostic reports
- the drain commits unconditionally when anything is staged, because the
inline path relies on finding the buffer empty (task 3 documents that)
- staging failure on the request path is now FATAL via wo_wal_stage_fatal:
the row is already in RAM and of the three verbs only insert could undo
itself, so continuing means RAM ahead of disk. One rule
- wal_die is now shared by both fatal points
Verified — the ack contract is the thing that could break, so it is what
was tested:
- just wovm-test: 36 suites (18 x both dispatch flavors) 0 fail, cli_smoke OK
- just db-bench-quick: 85 checks, 0 failures. The legs that matter:
crash.sN.0 — 612 acked rows all present after kill -9, which is the
BATCHING path (multi-shard requests, held replies, one barrier);
crash.s1.0 — 800 acked rows; restart.s1 and restart.sN replay byte-true
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
