writeonce/database/src/CODE-LOGIC.md

database/src — how the engine hangs together

The database engine is its own top-level directory, statically linked into every wovm and every runtime test binary (runtime/Makefile's DBSRC). One binary, unchanged. Format doc: docs/plan/oop-vm/04-db-binding.md. Memory-safety doctrine: the 9b design's section 6.

table.c — rows (iteration 9, Task 1)

VM values ──copy──▶ row slots (engine-owned malloc) ──copy──▶ fresh VM values
        wo_row_insert                              wo_row_read

• No VM pointer ever enters a slab; no slab pointer ever leaves. Encode copies per kind (Texts to db_text, owned objects flattened recursively to db_rec, containers element-wise); decode allocates fresh VM values from the caller's wo_rt. The GCREF kind is refused at encode — the compiler should have made that impossible (the GC bulkhead), the engine refuses it anyway.
• Rows never move. Slabs of 256 are malloc'd and kept for the table's life; the free-slot list recycles removed slots before any slab grows; the id hash maps id → slot. Ids are never reused (per-table counter, shard-interleaved S+1, S+1+N, …), which is also what makes the hash's tombstone sentinel safe.
• Storage is per-table since databasev2 2. @table(durable: false) sets WO_CLASSF_VOLATILE in the class descriptor (.wob v7), and db.c's table_is_durable gates all three mutation sites: a volatile table stages nothing, so it pays none of the fsync cost and is empty after a restart. Measured: 50 inserts wrote 1500 WAL bytes durable, 0 volatile. The three sites stayed three — the predicate is one function, not an inlined condition, precisely so this file's "nothing else may mutate storage" claim keeps holding.
• A mode mismatch refuses, it does not convert. If the log holds records for a class the loaded image now declares volatile, apply_record returns -2 (distinct from -1 corruption) and wo_wal_replay_ex reports the class id so main.c can name it. Silently skipping those records would resurrect nothing but would also hide a real migration; silently applying them would load rows into a table declared not to have any. wo_wal_replay remains as the NULL-out-param wrapper so the 156 WAL unit checks are untouched.
• Choke points: wo_row_insert / wo_row_remove carry the INDEX HOOK comments where Task 4's secondary indexes attach and Task 2's WAL stages its record. Nothing else may mutate storage.
• One deliberate file-static: g_classes for recursive frees (db_val_free has no context parameter). One process, one class table; revisit at iteration 8 (shards share the same immutable table).

wal.c — durability (iteration 9, Task 2)

The commit order IS the module: RAM apply → stage → one pwrite + one fdatasync → ack. wo_wal_commit returning 0 is the only thing "durable" means. Replay never touches the VM heap — payloads decode straight into engine-owned values and re-enter through the row API, so whatever hooks the choke points (indexes, Task 4) applies to replayed rows identically. Torn tails end the intact prefix and get overwritten by the next commit; CRC-valid-but-undecodable records fail replay loudly (corruption is not a tear). The crash battery in runtime/test/test_wal.c is the module's meaning proven: acked-over-a-pipe after commit, SIGKILL mid-stream, replay, zero acked-but-missing.

db.c — statement executors (iteration 9, Task 3)

One dispatcher, the builtin contract (0 ok, else WO_T_* + msg). The engine handles ride wo_rt.db / wo_rt.wal as opaque pointers set by main.c — NULL db traps WO_T_DB, NULL wal means RAM-only (the corpus's mode; WO_DATA opts into durability). Insert's contract: RAM apply through the row API, then stage + commit BEFORE returning — the builtin's return is the acknowledgment, so a failed commit un-applies the row and traps WO_T_IO rather than acknowledging what disk never got.

Verifying a change

• make -C runtime test — test_table is this directory's suite (round trips across kinds, nil encodings, shard interleave, slab growth, slot reuse, misuse), ASan+UBSan like every runtime test.
• just oop-e2e, just log-watcher — regression that linking the engine into wovm changed nothing observable (it is dead code until Task 3 wires the first builtin).

The slot-level surface (arc stage 3, 2026-08-21)

• Why it exists: the transparent DB actor executes a worker's statement on the owner shard, and VM heaps are never read cross-shard — so the requester encodes to engine slots on its own thread and the owner executes from slots, exactly the shape WAL replay already used.
• wo_db_val_encode exposes the in-gate for the RPC marshaler; wo_db_val_clone deep-copies an engine value (a get-field reply must outlive the row: a later serialized statement may free the slot); wo_row_insert_slots / wo_row_update_field_slot are the pre-encoded twins of insert/update (slot values consumed either way — installed on success, freed on failure); wo_db_exec_req (db.c) mirrors wo_builtin_db case for case with slot inputs and plain outputs, so a worker sees byte-identical traps and messages.
• The update refactor extracted row_apply_field_slot (the post-encode half: unique shadow-check, index fix-up, slot swap) shared by both entry points — the VM-value path's behavior is unchanged bit for bit.

The read-path index probe (2026-08-22)

• wo_idx_probe (table.c) answers a single-column equality from the index's hash buckets instead of walking slabs — the O(1) wiring the db-bench numbers demanded (reads were ~1.5k ops/s at p50 600µs on 20k rows; ~1.3M ops/s at p50 1µs after). idx_hash_key1 must reproduce idx_hash's single-column result bit for bit (same FNV over text bytes, same float canonicalization, same position mix) or probes and maintenance disagree on the bucket and rows silently vanish.
• The VERIFY step compares exactly as the slab walk compared (raw words for scalars/floats, byte equality for text; nil text == NULL bytes) — the hash canonicalizes only to FIND the bucket, so probe results are identical to scan results by construction.
• Composite indexes refuse (return 0) and callers keep the slab walk; both probe executors (wo_builtin_db and wo_db_exec_req) carry the same wiring, so worker shards get the speedup through the DB actor.
• The COMPILER half (emit.ml probe_key_of_where): a query whose where list contains var.col == key on a single-column-indexed column lowers its source to DB_PROBE; every where guard still runs over the candidates, so the guard — not the engine — stays the final arbiter. Keys are a plain identifier or an integer literal only; Float/Bytes columns excluded (engine raw-eq is narrower than VM float-eq, and a probe miss cannot be resurrected by a recheck). Pinned by tests/corpus/run/query-index-probe.

Group commit: one barrier per drain (databasev2 4 part A, 2026-08-28)

What changed: the engine used to commit per statement. db.c called wo_wal_commit immediately after every append, at all six sites, so each row change bought its own pwrite and its own fdatasync. Now the barrier belongs to the drain, not to the statement.

Where the barrier runs, and why there. A statement on a worker shard has no WAL to write — the runtime asserts workers hold neither db nor wal — so it marshals to shard 0 and parks. Shard 0 executes those requests in its envelope drain (wo_vm_adopt), and the drain now holds each reply instead of pushing it as the statement finishes. When the queue empties it issues one barrier, then releases every held reply.

Holding the reply is the whole mechanism. Pushing it early would unpark the requester before its record was durable; holding it means each writer is acknowledged after the barrier that carried its own record. That was always the intended contract — it was simply true by accident before, because every batch had exactly one member.

Why the queue is the boundary. Not a tick, and not a timer. A queue of one gives a batch of one, so a lone writer pays exactly what it paid before; the batch grows only when writes genuinely contend. A tick boundary would have added latency even with nothing to batch against, which is taxing an idle system to serve a busy one. There is nothing to tune, which is the point.

Why the inline path is asymmetric. A statement already on shard 0 stages and commits before returning, batch size one. It cannot hold a reply because there is nobody to reply to — it returns into its own fiber. Batching it would mean parking that fiber on the barrier, which is part B's machinery. Two consequences worth keeping in mind: single-shard configurations get no batching at all, by design; and the inline commit is only safe because the drain commits unconditionally whenever anything is staged, so the buffer is empty when an inline statement runs. If that ever stops holding, the inline path would make another statement's record durable early and acknowledge it to the wrong writer.

One rule for failure: once a statement has mutated RAM, the outcomes are durable or process death. It replaced three behaviours that disagreed — insert un-applied itself, while update and delete returned a catchable trap and left RAM ahead of disk, which their own comments said out loud. Batching would have multiplied that from one row to a whole batch. So a failed stage or a failed barrier now prints one diagnostic (operation, log path, errno, record count) and exits 3; WO_T_IO is unreachable from a write. Retrying is not offered because it is unsound: on Linux a failed fsync may already have discarded the dirty pages, so a second call can report success having written nothing. Replay is the recovery that works.

Measuring it. WO_WAL_STATS=1 makes the runtime print one line at exit — batches, records, peak batch, peak staged bytes. Opt-in, because it would otherwise pollute every durable program's output. The counters live in wo_wal rather than behind a builtin: they are diagnostic, not part of the language. db-bench's wmix N C leg exists to exercise this at all — mix writes on one op in ten with C=4, which produced a measured mean batch of 1.01, so it could never have shown whether batching worked.

If you are looking at this because writes got slower, check the mean batch first. Mean 1.0 means the mechanism is not engaging, which is expected for a serial writer or a single-shard configuration and a bug anywhere else.

Checkpoint: compaction by rewrite + rename (databasev2 3, 2026-08-29)

The problem: nothing ever removed superseded records, so the log grew forever and boot replayed all history. Measured before this: 20 000 rows seeded gave a 986 KB log; updating those same rows 20 000 times took it to 2.6 MB with the same live data.

Why one file and not a snapshot plus a tail. Postgres does the opposite — its WAL is a redo tail and the data lives in heap files, so a checkpoint flushes pages and then recycles log segments; it never compacts. It cannot: its records are page deltas, so a compacted redo log is not a store. Ours are full row images — apply_record implements UPDATE as remove-then-recreate — so a log of one record per live row is a complete store. That single difference deletes the control file, the redo pointer, the second recovery source and the separate process from this design. Recovery is not merely compatible with compaction; it is completely unaware of it.

Why rename is the whole crash-safety story. The dump goes to a temp file, which is fsynced, renamed over the live log, and then the parent directory is fsynced (the rename is atomic in-kernel, but the directory entry is not durable until the parent is — Postgres does the same for the same reason). Before the rename the live log is intact and the temp is not authoritative; after it the new log is complete. There is no instant at which a reader sees a mixture, so this needs no recovery logic of its own. What Postgres achieves with a redo pointer computed at checkpoint start and a control file written at the end, one syscall achieves here — because we can swap the entire data set atomically and Postgres cannot.

A crash mid-rewrite leaves a temp file. The next open removes it, and it is deleted rather than ignored because a file full of well-formed records sitting beside the log is exactly what a later reader mistakes for data.

Why the dump flushes periodically, and why it does NOT fsync when it does. stage() grows the staging buffer by doubling and never shrinks it, so pushing a whole store through one buffer would hold the entire store in RAM on top of the store — the unbounded growth databasev2 1 measured as how this engine dies. So the dump flushes every 256 records. It flushes with a plain write, not a commit: intermediate durability is worthless because the temp is not authoritative until the rename and is fsynced once immediately before it. Using the committing path cost one barrier per 256 records and made the pause 8× larger — measured 107 649 µs against 13 212 µs for a 2 MB live set, ~22 MB/s against ~181 MB/s.

Why the replacement is preallocated like the original. The WAL is preallocated so that appends never extend the file, which is what lets fdatasync alone serve as the ack barrier. A replacement opened without it would silently change that property, and the zero-padded tail the open-time scan relies on.

When it runs. Only where the staging buffer is empty — right after a barrier. Both write paths check: the drain (vm.c, after its commit and after releasing held replies, since those records are already durable and should not wait out a rewrite) and the inline path (db.c). Wiring only the drain left WO_SHARDS=1 never compacting, with its log growing forever: measured 536 KB where the multi-shard run held 446 KB.

The trigger compares the log against what the last compaction actually wrote, with an absolute floor. The denominator is measured rather than estimated, because estimating the live size means estimating Text and the compactor already knows the true number. There is deliberately no timer: Postgres needs one because its dirty buffers are not durable until flushed, and ours are durable at commit — an idle log does not grow.

A failed compaction is a missed optimisation, not a durability event. It leaves the original log intact and returns an error the callers ignore. It must never take wo_wal_commit_fatal's path, which exists for a different problem.

If you are here because a checkpoint misbehaved: WO_WAL_STATS=1 reports compaction count, the stop-the-world pause (max and total) and the last compaction's size. WO_CHECKPOINT_BYTES and WO_CHECKPOINT_RATIO move the policy; setting a tiny floor forces compaction in a few writes, which is how the gate tests it at all.

Keys-resident updates: read-modify-append, stage-here/commit-in-caller (databasev2 2/3, 2026-08-30)

The shape. A keys-resident row has no slab slot to mutate — its payload lives in the log — so row_apply_field_keys (table.c) does read-modify- append instead of a slot swap: borrow (folds the row's current value), append a WAL delta record (id, field, new value) chained off the row's current offset via a back-pointer, RAM-apply the index swap. wo_wal_fold_row_at is THE fold — written once, called by every reader (wo_row_borrow), by replay, and by compaction — so a read, a boot, and a checkpoint can never disagree about a chain's current value.

Stage-here, commit-in-caller — mirrors insert exactly. row_apply_field_keys stages the delta but does not commit and does not move the id map: table.c applies RAM and appends; db.c owns the barrier and the post-barrier map move, the same split insert already used (wo_wal_pend_drop / wo_db_flush_drops for insert; wo_wal_pend_repoint / wo_db_flush_drops for update). The caller captures the delta's own offset via wo_wal_next_offset() before calling in — insert's own koff pattern — since nothing between that capture and wo_wal_append_delta stages any other bytes on the WAL. back_off — the back-pointer a new delta chains from — checks a PENDING re-point (wo_wal_repoint_offset1) before falling back to the durable wo_row_offset1: two updates to the same row staged behind one drain's barrier must chain to each other, not both to the row's pre-drain offset, or the first update would be orphaned from the chain.

The unique shadow-check runs against a THROWAWAY buffer, never t->scratch. The row under update already occupies the table's one scratch buffer (wo_row_borrow refuses a nested borrow on the same table), so a candidate probe needs a buffer of its own — keys_fold_into, the fold-into-a-caller- supplied-buffer half of wo_row_borrow, bypasses the scratch gate for exactly this. A candidate updated earlier in the SAME uncommitted drain has its re-point only pending, so the candidate probe also consults wo_wal_repoint_offset1 — and wo_wal_fold_row_at itself reads the WAL's staging buffer (not yet durable) for an offset that falls inside it, so a same-drain candidate's NEW value is what a real @unique clash sees.

A keys-resident borrow holds ENGINE values, exactly wo_row_ptr's contract — restored 2026-08-30. table.h's opening doctrine: "the engine and the VM heap are two memory worlds crossed only by copy... a row stores NO VM pointer." keys_fold_into used to decode the fold's engine output to a VM value before handing the row back, which every OTHER reader of a borrowed row (db.c's GET_FIELD/PROBE, wo_row_read, and idx_hash/idx_cols_equal/ wo_idx_probe) was NOT written to expect — they all decode engine→VM themselves, on the assumption a borrow is engine-encoded like a slab row. Invisible for SCALAR/FLOAT (decode is identity either way), and un-exercised for TEXT/BYTES because the loader refused resident: keys outright until this task lifted it — nothing had ever read a keys-resident Text field through db.c at all. Fixed by making keys_fold_into stop decoding: the fold's engine output lands straight in the borrowed row's slots, wo_row_release frees them with db_val_free (not wo_drop_kind) exactly like table_destroy frees a slab row's fields, and row_apply_field_keys uses its already-engine-encoded nv directly instead of decoding a throwaway VM copy. No index function needed to change, and neither did db.c. Reproduced as a genuine ASan heap-buffer-overflow (a wo_str* read through the db_text* layout) before the fix, pinned by test_keys_resident_update_indexed_text (runtime/test/test_wal.c) after it.

Three limitations, shipped and documented rather than fixed:

1. Mid-drain stale reads. A request reading a row inside the same uncommitted drain as an earlier request's in-flight update to it may see the last durable value. Read-your-writes holds within a request, not across requests sharing a drain; closing it needs the fold to consult the staging buffer generally, not only for the same-drain unique shadow-check above.
2. Replay is O(N²) in a row's delta-chain length — apply_delta folds the pre-delta row, and wo_row_remove (called internally) folds the SAME offset again, so each replayed delta re-walks its whole chain.
3. Compaction triggers on byte ratio only — closed by databasev2 11: the fold reports hop count and row_apply_field_keys writes a full-row image (WO_WAL_UPDATE) past WO_DELTA_MAX_HOPS (16), so a hot row's chain is bounded in the update path itself; the checkpoint no longer carries that burden. wo_wal_should_compact also gained an absolute garbage term (WO_CKPT_ABS_BYTES).

Schema migrations (databasev2 12)

A @table class is the schema; the log is the database; boot compares them.

• The log describes itself. WO_WAL_SCHEMA (kind 5) is the head record of every fresh and every compacted log: per class its NAME, storage flags, and per field name + kind + the two encoding-relevant metadata words. Written lazily by stage() ahead of the FIRST real record — never for a log that stays empty, because durable: false programs have a documented zero-bytes contract. apply_record skips it before reading cid/id (its class count would be misread as a cid); replay does not count it.
• The diff is name-keyed (wo_schema_diff). Classes match by name, fields by name + kind, owned references (fclass) by the NAME the number resolves to — so pure declaration reordering costs only a cid remap, which closes the old silent hole where reordering decoded rows into the wrong class. Verdicts are per-class POISONS carried in the plan: retype, same-shape delete+add (a disguised rename), vanished class, storage-flag change, and the embed closure (any class whose stored values carry a CHANGED class's old sub-shape, to a fixpoint). A poison forces the transcode and bites only when a record of the class is actually met — no rows, no verdict.
• The migration is a record-level transcode (wo_wal_migrate), not a replay: no id maps, no indexes, no keys-resident logic. Old shapes decode through a classdesc shim built from the stored schema; embedded cids are renumbered by mig_fixup_cids (owned values carry a cid on the wire); surviving fields move slots, deleted values are freed, added fields take enc_val(0) — the kind's zero. Delta back-pointers rewrite through an offset map, and a delta on a deleted field is SPLICED: it maps to its own target, so later deltas step over it. Temp + fsync + rename, compaction's own crash discipline — a kill anywhere leaves the old log authoritative, including a kill after the temp is complete (test_migrate_crash_before_rename).
• Legacy logs (no head record) replay exactly as before and adopt the head at their next compaction. v1 verbs are add and delete only; rename wants @renamed_from (v2), data/seed migrations are v2.