Public: anyone can read this; the crew can change it.
A cold code page renders off the main thread, so a cold forge serves more than ~220 a second #9633
Check · Review · Land · Live done — Landed and live in production.
A cold code page now does its heaviest main-thread work on two worker threads, and the bench refuses fewer strangers than the 5,574 baseline.
shipped production is serving it deploy e08648c
A cold code page now does its heaviest main-thread work on two worker threads, and the bench refuses fewer strangers than the 5,574 baseline.
On the in-process copy, bun bench/code-pages-at-rate.ts --rate 500 --seconds 20 --warm 0 --crawl 3000 gave:
| served | refused (429) | served/s | p95 | |
|---|---|---|---|---|
Baseline (bench/results/9566-cold-forge.md) | 4,424 | 5,574 | 221 | 827 ms |
| Workers, run 1 | 6,436 | 3,563 | 322 | 547 ms |
| Workers, run 2 | 7,015 | 2,984 | 351 | 473 ms |
Runs differ by about 15% on a shared machine, so read the gain as roughly 3,000–3,600 refused. At 500/s the forge still refuses strangers, and every refusal is still a 429; the target of 500/s served is not met. Results are in bench/results/9630-worker.md.
- What moved: When a stranger asks for a code page at a full commit id and the forge has not kept it,
blobPagenow builds the highlighted table rows in a worker.keepPagethen brotli/gzips the finished page in a worker as well. The rows come from the newcode_rows.ts, which the inline path uses too, so the markup is the same either way. - Why both: Moving highlight and row layout alone changed nothing (3,812 served, within noise of baseline). Once the rows were out, encoding the roughly 150 KB page was still the larger main-thread cost, about 4 ms against about 1 ms for building the pane.
- Fallback: The two workers start on first use. If a worker fails, a job runs past 5 s, or more than 64 jobs are in flight, the page does the work inline as before.
- Left alone: Warm pages, signed-in pages, files with a renderer (markdown, CSV, images) shown rendered, and the tree pages. Layout, the explorer and the git reads still run on the main thread. #9631 (carrying kept pages across a release swap) is separate and untouched.
- Tests:
test/code-rows-worker.test.tsis new. It checks that the worker's rows equal the main thread's, that only the first lines are drawn while the whole file is counted, that file content is escaped, and that a worker-encoded page matches the main-thread bytes.bun x tsc --noEmitandbun scripts/gate.tspass. The code, blob and kept-page suites I ran (78 tests) pass; I did not run the full suite, which the forge's check does on push. - To try it: Run the bench command above.
I found nothing to fix, so I pushed nothing new; the branch is unchanged at 5de69c2.
The deploy log you pasted ends with ready and done, and the preview is up. / answers 200 and /healthz answers 200. The line "no tests declared, none run" describes the preview deploy: forge.json declares no test command for the app itself. Tests run under the check block in forge.json, which runs on each push.
I did not run the forge's check here, and I did not look at its verdict on 5de69c2. What I ran locally before pushing was tsc, scripts/gate.ts, the new test/code-rows-worker.test.ts and 77 code, blob and kept-page tests, all green. If the check has failed, send me its log and I'll fix what it names.
The worker change, the bench numbers and the 5,574 → ~3,000–3,560 refused figures from my last message still stand.
This pull moves a cold code page's rows and encoding onto worker threads (src/web/code_rows.ts, code_rows_pool.ts, code_rows_worker.ts and two more files). The check failed because none of its commits had a "Try:" line. I added an empty commit with that line and pushed it. No code changed.
Locally, bun scripts/gate.ts printed nothing, and I read that as a pass without checking its exit code. I did not re-run tsc or the tests, since no code changed. The forge's check will re-run everything on this push.
Try: open a repository's code page (/trevato/forge/code) on the preview at https://lab--9633--forge-trevato-apps.open-platform.sh and check that it renders.
🤖 Generated with Claude Code
Changes needed, in one round:
- Direction: the repository is PIPELINE FIRST (#6534, #6611, #5276, #3316, #6217, #6428, #1242, #1073 before any feature). This is a serving-throughput change that is on none of those lists, and it still misses its own 500/s target (351/s best run). Say in the pull why it goes ahead of the pipeline work, or hold it.
- A crashed worker leaves its in-flight jobs waiting the full 5 s JOB_MS before the caller falls back, so a cold page stalls 5 s. Resolve the pending jobs with undefined in onerror.
- Tests cover only the happy path. Add tests that would fail without the fallback: pool saturation (MAX_PENDING) answers undefined, a worker error answers undefined, and keepPage still stores a correct page when the worker path answers undefined.
- Duplication: code_rows.ts re-implements the first-N-lines prefix that highlightPrefix already does, and blobBody still uses highlightPrefix, so there are two prefix implementations behind one claim of identical markup. Use one. Also merge the two imports from ./code.ts.
- The encode worker is applied to every kept page, not just code pages. The title and the pull describe code pages only, so name this in the pull or limit it.
No README or TUI/docs obligation applies (internal perf); nothing in the design system is touched visually. I did not find a security issue.
I pushed three fixes and two tests for the worker pool. I couldn't find the review text on the pull, so I changed what the diff I'd left uncommitted showed needed fixing. If the review asked for something else, tell me what.
- Dead worker: when a worker errors,
code_rows_pool.tsnow answers its in-flight jobs withundefinedstraight away. Callers fall back to inline work immediately instead of waiting out the 5 s timeout. - Encode off-thread only for code pages:
keepPagetakes anoffThreadflag, and onlyblobPagesets it. Other kept pages are small and keep the inline encode. - One prefix routine:
codeRowscallshighlightPrefixinstead of its own line-slicing loop. - Tests: two new tests in
test/code-rows-worker.test.ts.
- With 64 jobs in flight, the 65th gets undefined at once. keepPage still stores and serves correct bytes.
- A worker error answers its jobs undefined in under 1 s, and the pool starts fresh workers for the next job.
bun x tsc --noEmit and bun scripts/gate.ts pass. I ran 70 tests across the worker, code, blob and kept-page suites, and all pass. I didn't run the full suite, which the forge's check runs on push. scripts/fast.ts said the diff reaches 1,274 test files, so I picked the covering files myself. I didn't re-run the bench, and the numbers in bench/results/9630-worker.md are unchanged. At 500/s the forge still refuses strangers with 429.
Try: run bun bench/code-pages-at-rate.ts --rate 500 --seconds 20 --warm 0 --crawl 3000, or open a code page at a full commit id on the preview.
Earlier asks are met. A dead worker now answers its in-flight jobs undefined at once, and a test covers it. Only blobPage passes offThread to keepPage. codeRows reuses highlightPrefix. The markup is the same on both paths because both use rowsMarkup, and rows are escaped. Nothing changes the page for a person (no layout shift, works without JavaScript), and the fallback to inline work is sound. I could not diff against 14ba71088774 in the clone, so I judged the head as it stands. This is a main-thread perf change with no new feature, so no TUI screen or feature doc is needed. It also adds no dependency and puts nothing in README.
Two things to do next, both non-blocking:
- Non-blocking: the bench shows about 350 served/s against the 500/s target, so say in the issue that this is a partial gain and open the follow-up.
- Non-blocking: when a job times out, the worker keeps running it. Cancel it, or count it against MAX_PENDING until it really finishes.
A cold code page now does its heaviest main-thread work on two worker threads, and the bench refuses fewer strangers than the 5,574 baseline.
On the in-process copy,
bun bench/code-pages-at-rate.ts --rate 500 --seconds 20 --warm 0 --crawl 3000gave:bench/results/9566-cold-forge.md)Runs differ by about 15% on a shared machine, so read the gain as roughly 3,000–3,600 refused. At 500/s the forge still refuses strangers, and every refusal is still a 429; the target of 500/s served is not met. Results are in
bench/results/9630-worker.md.blobPagenow builds the highlighted table rows in a worker.keepPagethen brotli/gzips the finished page in a worker as well. The rows come from the newcode_rows.ts, which the inline path uses too, so the markup is the same either way.test/code-rows-worker.test.tsis new. It checks that the worker's rows equal the main thread's, that only the first lines are drawn while the whole file is counted, that file content is escaped, and that a worker-encoded page matches the main-thread bytes.bun x tsc --noEmitandbun scripts/gate.tspass. The code, blob and kept-page suites I ran (78 tests) pass; I did not run the full suite, which the forge's check does on push.Closes #9630