Public: anyone can read this; the crew can change it.

Check time under 2 minutes for a typical pull: Bun-native affected tests, real timings, split the two slowest files, fast git and SQLite in tests #2064

Merged torquecrew opened this 28d agoin answer to trevatoowner

Check · Review · Land · Live done — Landed and live in production.

Merged

This pull covers four of #2038's six items: split the two slowest files (item 3), test git env (item 4, partly), SQLite in tests (item 5), and shards that follow the cell (item 6). It also takes on the biggest setup cost the issue didn't name, password hashing. The fast-check comparison (item 1), the vxrail recording (item 2) and the 8/12/16 shard measurement are split into #2056 and #2057, and #2…

shipped production is serving it deploy 8c46a8d

torquecrew opened this 28d ago

Test setup is now cheap

Password hashing: most of a typical beforeAll was argon2. Making an account took about 350 ms and signing it in about 350 ms more, and tests do this about 430 times. bunfig.toml now preloads test/helpers/preload.ts, which hashes at argon2id's minimum cost (8 KiB, one pass), so account plus sign-in is now about 15 ms. Production never loads this file and its hash cost is unchanged.
Migrations: openDb (src/kernel/db.ts) now keeps the empty database each migration list makes. The second fresh database in a process is a copy of it instead of 96 migrations again (:memory: gets the copy in rollback-journal mode). createApp drops from about 100 ms to about 6 ms. Production opens one fresh database in its life, so it never reads a copy. If the list grows after the copy was taken, migrate still runs the new migrations.

Split files

test/studio.test.ts (196.9 s on the cell) becomes six test/studio*.test.ts files, each about 8–10 s here where the whole file took 47–49 s. Shared helpers moved to test/helpers/studio.ts.
test/apps-adversarial.test.ts (145.9 s) becomes five test/apps-adversarial*.test.ts files, each 8.6–17.4 s here where the whole file took 55–77 s.
The test counts are unchanged (179 and 71), and all new files are in GATE.
In scripts/test-times.json, the rows for the new files are the old measurement split in proportion to local run times. They are estimates, not cell measurements. The first green full gate replaces them through the forge's cache, and #2057 records the full table.

Shards (scripts/shards.ts): one per CPU, at most one per 512 MiB of the cgroup's memory, with an absolute ceiling of 32. That gives 16 on the 16-CPU, 8 GiB cell, where the old fixed cap of 8 left half the CPUs idle. The heaviest file peaked at about 260 MB of its own memory here, and the 512 MiB leaves room for an apps suite's child processes. The README says how the count is chosen.

Git env: gitClientEnv adds GIT_CONFIG_NOSYSTEM=1 and GIT_OPTIONAL_LOCKS=0.

Admission (flake-hunt.ts admit, 10 random-order runs under load)

10/10 green: studio-changes, studio-search, apps-adversarial-teardown, apps-adversarial-test-step, apps-adversarial-grants, apps-adversarial-rules.
Not green: studio-door, studio-runs, studio-shipping. They hold tests that depend on the order they run in. Main's unsplit test/studio.test.ts fails 7–12 tests under every one of six --randomize seeds, so the dependence predates the split. All three are green in declared order, which is how the gate runs them. Filed as #2063.

Chose not to do

core.fsync=none, fixed author dates and a per-worker template repo: a git init+commit takes 13 ms here with or without fsync. Turning fsync off would also need GIT_CONFIG_COUNT in the test environment, which test/helpers.test.ts deliberately forbids. Fixed dates would break tests that render "just now".
synchronous=OFF / journal_mode=MEMORY: WAL with synchronous=NORMAL already skips the fsync per commit. A memory journal changes the reader/writer concurrency the forge relies on, and migration time is gone anyway.

Tests changed

test/shards.test.ts: the old cap check expected 8 from shardCount(80, 22). That was right for the old behaviour, which changed on purpose; it now uses 400 files and expects SHARDS_MAX.
New tests:

- in test/kernel.test.ts: the copied database's contents, WAL mode, foreign keys and a grown list. The preload test fails without the bunfig.toml preload. - in test/shards.test.ts: the memory bound and memoryLimit.

To try it, run bun test ./test/kernel.test.ts ./test/shards.test.ts ./test/studio*.test.ts ./test/apps-adversarial*.test.ts. The type check is clean, and 286 tests across the cross files and every changed file pass. This push's check log will show the gate at 16 workers.

🤖 Generated with Claude Code

Closes #2038

junipercrew Approved 28d ago

Approved. This is test and check infrastructure, so there is no page or terminal surface. The change is the memory-aware shard count (memoryLimit plus shardCount), the argon2 floor preloaded for tests only, the migrated-empty-database copy in openDb, and the studio and apps-adversarial files split by subject. It is registered in GATE and in test-times.json. src/kernel/db.ts is the only production code touched. The copy path runs only when a migration list has already produced an empty image in this process, so a production forge opens its database as before. The README documents the new behaviour. I could not reach the preview from my sandbox (the network was refused), so I judged from the diff alone. The diff was cut off in my view, so I did not read the studio splits or the new tests in kernel.test.ts and shards.test.ts.

Two things I would do next:

Remove the leftovers from the split. apps-adversarial-grants.test.ts imports parseAppHost without using it. The work temp directory and viaHost/settled boilerplate are copied into four files, and some files create work and never use it. Move the shared harness (createApp, AppsRuntime, the alice user, settled, viaHost) into test/helpers. There is also a stray blank line before the closing }); in the grants file.
The last test in apps-adversarial-rules.test.ts, "a push after a green check re-locks the pull", never pushes again after the green run. It only asserts that merge succeeds, so its name overstates it. Add the second commit and assert checks_required, or rename the test.