Repository navigation
perf: WAL fsync-per-write is a full fsync() with no batching/relaxed-durability escape hatch #302
Description
Context
Found while diagnosing why temporal-reasoning's git-history ingestion pipeline was blocking its host process's event loop. Every commit ingested there issues on the order of 20-25 separate db.execute()/checkpoint() calls (one per fact for :contains/:depends-on edges, per the workaround for #287), and each one turns into a full, synchronous fsync in WalWriter::append_entry (src/wal.rs:153-158):
pub fn append_entry(&mut self, tx_count: u64, facts: &[Fact]) -> Result<()> {
let entry_bytes = serialize_entry(tx_count, facts)?;
self.file.write_all(&entry_bytes)?;
self.file.sync_all()?;
Ok(())
}For a bulk-load / high-write-volume caller (ingestion, migration, seeding), this is the dominant cost — dozens of synchronous disk round-trips per logical unit of work, each one a full fsync.
The consuming application already worked around the blocking symptom by moving these calls off its async event-loop thread onto a dedicated worker (confirmed safe: minigraf-ffi's UniFFI/ctypes bindings release the GIL for the duration of the call, and Arc<Mutex<Minigraf>> in minigraf-ffi/src/lib.rs serializes cross-thread access correctly). That fixes unresponsiveness but does nothing about the raw fsync volume/cost, which is a minigraf-side concern.
Proposed improvements (independent, increasing risk/leverage)
-
sync_data()instead ofsync_all()—File::sync_data()(fdatasync) skips flushing inode metadata (mtimes etc.) thatsync_all()(fsync) also flushes, while still guaranteeing the appended data + the file-length metadata needed to read it back are durable. Since the WAL is pure append, there's no metadata caller here depends on beyond size. Low risk, no format change, should be a straight swap inwal.rs:156. -
Let non-colliding facts share one
commit()/one fsync today —WriteTransaction::commit()(db.rs:1012-1059) already does exactly one WAL write for however many facts were staged viaexecute(), since it re-stamps everything with a singletx_countbefore applying. That's unrelated to the bug: EavtKey and AevtKey missing value bytes causes silent data loss when batching multi-valued facts #287 bug (which only bites when multiple facts share entity+attribute+valid_from+valid_to). So call sites that today issue one baredb.execute()per fact for unrelated entities/attributes (e.g. module-level facts + parent-commit edges + watermark update in one ingestion step) could switch tobegin_write()+ Nexecute()+ onecommit()and collapse N fsyncs into 1, with zero format change and no interaction with bug: EavtKey and AevtKey missing value bytes causes silent data loss when batching multi-valued facts #287. Might be worth calling out explicitly in the docs/README as the recommended pattern for bulk writers, since it's easy to miss thatcommit()already batches. -
Opt-in relaxed durability mode (biggest lever, real tradeoff) — expose something like SQLite's
PRAGMA synchronous=NORMAL: skip the per-write fsync and rely on the OS page cache + a periodic/explicitcheckpoint()(which already does a heavier flush) for durability. Risk window is an OS crash/power loss between writes and the next checkpoint (not an ordinary process crash — the WAL bytes and the extended file length are still in the OS's write-back cache and become durable on next fsync regardless of which process wrote them). For bulk-loaders that can safely re-run from a watermark/checkpoint on failure (like git ingestion), this could be a large throughput win. Would need to be an explicit, well-documented opt-in givenMinigraf's current default is "every transact/retract is fsync-durable immediately."
Suggested priority
(1) is a near-zero-risk quick win. (2) is mostly a docs/call-site change, not a library change — maybe just needs clearer documentation. (3) is the real design decision and probably deserves its own issue/RFC if there's interest — flagging it here for visibility since it fell out of the same investigation.
Happy to provide more detail on the ingestion-side call pattern that motivated this if useful.
Activity
perf: offload ingestion DB writes to a dedicated thread to unblock th…
perf: WAL fsync-per-write batchin 7AE1 g/relaxed-durability escape hatch (#302
adityamukho commentedon Aug 25, 2026
Fixed by #353 (squash commit e2f4912). Full writeup in the PR description and
docs/superpowers/specs/2026-08-24-wal-sync-mode-design.md.