writeonce/database/src/wal.h
shoney.arickathil 40d56c4664 perf(wal): checkpoint measured — 2.16x space, 1.78x boot, 2.7ms pause — T5
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>
2026-08-29 06:20:13 +02:00

203 lines
10 KiB
C

/* wal.h — typed-row write-ahead log + boot replay (iteration 9, Task 2).
*
* The c-runtime plan's shipped pattern (phases D/E), generalized to typed
* rows. The commit order is doctrine, verbatim:
*
* RAM apply → wal_append (staged) → wal_commit (write + fdatasync)
* → only then is the write ACKNOWLEDGED
*
* Record framing — replay-whole-or-not-at-all:
*
* record := len u32 | crc u32 | payload | mark u32
* len = payload byte count (never 0; 0 = preallocated tail, stop)
* crc = CRC32 of payload
* mark = 0x574F4C31 "WOL1" — written LAST, so a record without its
* mark is torn by definition
* payload := kind u8 | class_id u32 | row_id u64 | body
* kind : 1 insert (body = the row's fields, engine encoding below)
* 2 remove (no body)
* 3 update (reserved for Task 5)
*
* Field encoding in a body walks the class table's kinds:
* SCALAR 8 bytes
* TEXT u32 len | bytes (0xFFFFFFFF = nil)
* OWNED u8 0 = nil, or u8 1 | u32 class_id | fields recursively
* MULTI u8 0 = nil, or u8 1 | u8 elem_kind | u32 len | elements
* MAP u8 0 = nil, or u8 1 | u8 kk | u8 vk | u32 len | k v pairs
*
* Replay decodes payloads STRAIGHT into engine-owned values — the VM heap
* is never involved (boot must not depend on a VM existing yet), and rows
* re-enter through the same choke-point row API, so Task 4's indexes are
* rebuilt for free. A torn tail (short record, bad CRC, missing mark) drops
* everything from the tear onward — never a partial record, never a record
* after a tear. Little-endian on-disk, matching the .wob loader's platform
* note.
*
* wo_wal_check is the offline oracle the crash battery verifies with: it
* walks a WAL file with no engine at all and reports how many records are
* intact and where the intact prefix ends. */
#ifndef WO_WAL_H
#define WO_WAL_H
#include "table.h"
#define WO_WAL_MARK 0x574F4C31u /* "WOL1" LE */
enum { WO_WAL_INSERT = 1, WO_WAL_REMOVE = 2, WO_WAL_UPDATE = 3 };
typedef struct wo_wal {
int fd;
/* databasev2 4: where this WAL lives, so a durability failure can name
* the file it could not write. An abort diagnostic without the path
* sends an operator hunting. Owned here, freed by wo_wal_close. */
char *path;
uint64_t off; /* next write offset (the intact tail) */
/* staged batch: appended by wal_append_*, flushed by wal_commit */
uint8_t *buf;
size_t len, cap;
/* databasev2 4: group-commit diagnostics. Batching is worthless if
* batches are always one, and a throughput change would then have come
* from somewhere else — so the mechanism is measured, not assumed.
* peak_staged also settles whether the batch needs a cap with a number
* instead of a guess. Reported at exit under WO_WAL_STATS. */
uint64_t stat_batches; /* non-empty commits */
uint64_t stat_records; /* records those commits carried */
uint64_t stat_peak_batch; /* most records in one barrier */
uint64_t stat_peak_staged; /* most bytes staged behind one barrier */
/* databasev2 3: bytes the last compaction wrote. The trigger compares the
* log against THIS rather than an estimate of the live set — estimating
* would mean estimating Text, and the compactor knows the true number. */
uint64_t compacted_bytes;
/* databasev2 3: the preallocation this log was opened with. Compaction
* MUST give the replacement the same one: the WAL is preallocated so that
* appends never extend the file, which is what lets fdatasync alone be the
* ack barrier. A replacement without it silently weakens durability. */
uint64_t prealloc;
/* databasev2 3: what compaction actually did, reported under WO_WAL_STATS.
* The PAUSE is the number the spec refused to assume — compaction is
* stop-the-world, so its duration is the cost being weighed. */
uint64_t stat_compactions;
uint64_t stat_compact_us_max;
uint64_t stat_compact_us_total;
} wo_wal;
/* Open (create if missing) and preallocate [prealloc] bytes (best-effort;
* a filesystem without fallocate still works). Positions the write offset
* at the end of the INTACT record prefix — an existing file is scanned the
* same way replay scans it, so a torn tail is overwritten, not appended
* after. 0 ok, -1 errno-style failure. */
int wo_wal_open(wo_wal *w, const char *path, uint64_t prealloc);
void wo_wal_close(wo_wal *w);
/* Stage a record for the row that MUST already be applied to RAM (the
* commit-order doctrine). Insert/update read the row via wo_row_ptr.
* 0 ok, -1 OOM / no such row. */
int wo_wal_append_insert(wo_wal *w, wo_db *db, uint32_t class_id, uint64_t id);
int wo_wal_append_remove(wo_wal *w, uint32_t class_id, uint64_t id);
/* UPDATE re-logs the whole row (KISS: replay replaces — remove + re-create
* with the same id; the prefix/suffix delta trick from the survey is a
* later optimization, recorded). Call AFTER the RAM update. */
int wo_wal_append_update(wo_wal *w, wo_db *db, uint32_t class_id, uint64_t id);
/* databasev2 4: which half of the barrier failed. A pwrite failure and an
* fdatasync failure are different operational problems (a short write vs a
* device refusing the flush), so the diagnostic must name the right one. */
#define WO_WAL_ERR_WRITE (-1)
#define WO_WAL_ERR_SYNC (-2)
/* The process exit status for a durability failure.
*
* 74 is sysexits' EX_IOERR, chosen deliberately over a small number: 1 is a
* trap and 2 is a loader refusal, but 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 is supposed to help
* diagnose. The low range belongs to programs; the runtime takes a high one. */
#define WO_EXIT_DURABILITY 74
/* Write the staged batch and fdatasync — the ack line. Empty batch = ok,
* no syscall. 0 ok, WO_WAL_ERR_WRITE / WO_WAL_ERR_SYNC on failure (the
* batch stays staged: a failed commit consumes nothing). */
int wo_wal_commit(wo_wal *w);
/* databasev2 3: the checkpoint trigger, as a PURE decision so it can be tested
* without a store — which is the only way a policy like this gets tested at all.
*
* [used] the log's used bytes; [last] what the LAST compaction wrote (0 if it
* has never run); [floor] the size below which compacting is not worth it;
* [ratio] the multiple of [last] that counts as too much history.
*
* The denominator is the last compaction's MEASURED output rather than an
* estimate of the live set: estimating would mean estimating Text, and the
* compactor already knows the true number.
*
* There is deliberately NO TIME component. Postgres' CheckPointTimeout exists
* to bound data loss from unflushed buffers; our records are durable at commit,
* so a checkpoint only reclaims space and shortens boot. An idle log does not
* grow, so a timer would fire with nothing to do.
*
* 1 = compact now, 0 = leave it. */
int wo_wal_should_compact(uint64_t used, uint64_t last, uint64_t floor, uint32_t ratio);
/* Defaults, overridable at boot by WO_CHECKPOINT_BYTES / WO_CHECKPOINT_RATIO.
* The knobs are what make the policy testable: a test sets a tiny floor and
* forces compaction in a few writes instead of waiting for megabytes. */
extern uint64_t wo_wal_ckpt_floor;
extern uint32_t wo_wal_ckpt_ratio;
/* databasev2 3: the temporary file compaction writes before the swap. Named
* next to the log so it lands on the same filesystem — rename(2) is only
* atomic within one. Boot removes a stale one (a crash before the rename). */
#define WO_WAL_TMP_SUFFIX ".compact"
/* databasev2 3: rewrite the log as one INSERT record per LIVE row, then swap
* it in with rename(2).
*
* Recovery is deliberately untouched: the result is an ordinary log in the
* ordinary grammar, replayed from byte 0. Crash safety comes from rename being
* atomic — before it the live log is intact and the temp file is not
* authoritative; after it the new log is complete. There is no window in which
* a reader sees a mixture, so this needs no recovery logic of its own.
*
* REFUSES if anything is staged (returns -1 without touching the log): those
* records would be written into a file about to be replaced. Callers must
* invoke this only where the staging buffer is empty — right after a barrier.
*
* A failure is a MISSED OPTIMISATION, not a durability event: the original log
* is left usable and the process keeps running. It must not take the fatal
* path wo_wal_commit_fatal takes.
*
* 0 ok, -1 on any failure. */
int wo_wal_compact(wo_wal *w, wo_db *db);
/* databasev2 4: a record could not even be STAGED (the row is already in
* RAM, so this is the same unrecoverable position as a failed barrier — see
* wo_wal_commit_fatal). Never returns. */
void wo_wal_stage_fatal(const wo_wal *w);
/* databasev2 4: commit, or END THE PROCESS.
*
* The one rule this iteration introduces: once a statement has mutated RAM,
* the only outcomes are durable or process death. Retrying is not an
* alternative — on Linux a failed fsync may already have discarded the dirty
* pages, so a second call can report success having written nothing. The
* recovery that works is replay, which returns the last durable state.
*
* [nrec] is the number of records in the batch, for the diagnostic only.
* Returns on success; never returns on failure. */
void wo_wal_commit_fatal(wo_wal *w, uint32_t nrec);
/* Boot replay: apply every intact record to [db] in order. Ids re-enter
* exactly as logged; each table's next_id advances past the replayed ids
* that belong to this shard. Returns the number of records applied, or -1
* on open failure / a record naming an unknown class (corruption beyond
* what a torn tail explains). A torn tail is NOT an error: replay applies
* the intact prefix and reports it. */
int64_t wo_wal_replay(const char *path, wo_db *db);
/* Offline verification (no engine): scan [path], count intact records.
* *intact_bytes (optional) = where the intact prefix ends. -1 = open
* failure. */
int64_t wo_wal_check(const char *path, uint64_t *intact_bytes);
#endif /* WO_WAL_H */