[magurotuna]

Motivation

Every cache commit in vp run waits for an fsync of the WAL. The cache opens the database in WAL mode but leaves SQLite's default synchronous=FULL, and saving one task's result commits twice: once for cache_entries and once for task_fingerprints. Each cache miss therefore adds two disk syncs before the task counts as finished, so the tasks that depend on it start later.

Each run in the chart starts right after vt cache clean, so every task misses the cache and saves its result. The two workspaces are synthetic pnpm workspaces. Each package has one cached task, build, which is a Node.js script that joins five one-line files into dist/out.js. The 3 packages form a chain (app → lib → utils). Each of the 100 packages depends on 0–3 other packages, chosen by a seeded random generator.

Notice that macOS gains almost nothing with this change. This is because a sync there is already fast; SQLite calls fsync(), and on macOS, fsync() moves the data to the drive, but the drive can keep it in its own cache and write it later (fsync(2)).

I haven't measured Windows.

Changes

• Add the enable_wal function, which runs PRAGMA journal_mode=WAL and checks the returned mode. It sets synchronous=NORMAL only if the result is wal. With NORMAL, commits no longer sync; SQLite syncs the WAL at checkpoints, including the checkpoint when the last connection closes.

Safety

• In WAL mode, synchronous=NORMAL can't corrupt the database after a power loss or an OS crash, but it can roll back the most recent commits (SQLite docs). For this cache, a rolled-back commit means a cache miss on the next run, so the task runs again. The cache exists only to save time, so this is acceptable.
  • In one rare case, the next run fails instead. When a new entry replaces an old one, vp run first deletes the old entry's output archive. If the new entry is rolled back, the old entry points to the deleted archive, and the run fails with "Cache restore failed. Run vp cache clean to clear the cache". This can already happen on main, because nothing syncs the archives.
• PRAGMA journal_mode=WAL does not report an error when SQLite cannot use WAL. It keeps the current mode, usually a rollback journal, and returns that mode's name. With a rollback journal, NORMAL has a small chance of corrupting the database after a power loss, so synchronous stays FULL unless the result is wal.
• A process crash, including SIGKILL, does not lose committed transactions, because the data is already in the OS page cache.
• The schema and the file format don't change, and SQLite doesn't store synchronous in the file. I checked that main and this PR can use the cache entries that the other wrote.

[wan9chi]

Nice!