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

Bun 1.4.4 leaves two tests red on main: the keep-alive handover and a websocket handshake the app refuses #12728

Merged opened this

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

Merged

Both tests are green on Bun 1.4.4. The keep-alive test was reproduced red and now passes. The websocket test never failed here, so its fix is a precaution.

Notes

shipped production is serving it deploy e5f3b40

opened this

Both tests are green on Bun 1.4.4. The keep-alive test was reproduced red and now passes. The websocket test never failed here, so its fix is a precaution. I ran those two test files, tsc --noEmit and bun scripts/gate.ts (clean); I did not run the full suite.

Keep-alive handover (test/release-handover-keepalive.test.ts): Bun 1.4.4 closes an idle kept-alive connection itself when stop(false) runs. So the old forge has nothing left to answer on a warmed connection. That is why the next fetch got 404 from the release.

  • The test now warms one raw TCP connection (openRaw, a small HTTP/1.1 client over Bun.connect) so reuse is provable.
  • It then asserts the handover closes that connection, and that fresh connections reach the release (404).
  • The connection: close wrapper in src/http/server.ts stays. It still covers a connection that is busy when the port is handed over, but no test exercises that case now.

Websocket refusal (connectUpstream in src/apps/proxy.ts): test/apps-ws.test.ts passed here, alone and with the whole file. I couldn't reproduce the 502 "app stopped answering" page.

  • My guess is a race: the close event arriving before the error event that says the app refused the handshake.
  • The close handler now resolves a tick later (setTimeout(…, 0)), so the refusal flag is set before it is read.
  • This is unconfirmed. If the test still fails in the forge's check, the cause is elsewhere.

I didn't touch the tests tracked in #12716 and #12618.

To try it: bun test test/release-handover-keepalive.test.ts test/apps-ws.test.ts.

🤖 Generated with Claude Code

Closes #12727

2 earlier comments · 1 event
commented

Both tests are green on Bun 1.4.4. The keep-alive test was reproduced red and now passes. The websocket test never failed here, so its fix is a precaution. I ran those two test files, tsc --noEmit and bun scripts/gate.ts (clean); I did not run the full suite.

Keep-alive handover (test/release-handover-keepalive.test.ts): Bun 1.4.4 closes an idle kept-alive connection itself when stop(false) runs. So the old forge has nothing left to answer on a warmed connection. That is why the next fetch got 404 from the release.

  • The test now warms one raw TCP connection (openRaw, a small HTTP/1.1 client over Bun.connect) so reuse is provable.
  • It then asserts the handover closes that connection, and that fresh connections reach the release (404).
  • The connection: close wrapper in src/http/server.ts stays. It still covers a connection that is busy when the port is handed over, but no test exercises that case now.

Websocket refusal (connectUpstream in src/apps/proxy.ts): test/apps-ws.test.ts passed here, alone and with the whole file. I couldn't reproduce the 502 "app stopped answering" page.

  • My guess is a race: the close event arriving before the error event that says the app refused the handshake.
  • The close handler now resolves a tick later (setTimeout(…, 0)), so the refusal flag is set before it is read.
  • This is unconfirmed. If the test still fails in the forge's check, the cause is elsewhere.

I didn't touch the tests tracked in #12716 and #12618.

To try it: bun test test/release-handover-keepalive.test.ts test/apps-ws.test.ts.

🤖 Generated with Claude Code

commented

The only red this pull caused was timers-named, and that is fixed and pushed. I ran test/timers-named.test.ts and test/release-handover-keepalive.test.ts locally (both pass) and bun scripts/gate.ts (clean). I did not run tsc or the full suite this round.

The cause was the setTimeout I added in connectUpstream (src/apps/proxy.ts), which makes the refused-handshake verdict wait one tick after the close event. The test requires every server-side timer to run inside during or to be on its allowlist. I added a one-line entry in test/timers-named.test.ts naming that timer, because the callback just resolves a promise and has no work to wrap in during.

An earlier push carried the // real-timer: comment on the raw-socket wait in test/release-handover-keepalive.test.ts. The gate passes now.

The other red tests in the check log — check-install, apps-open-host and ask-to-reach — the check log says are red on main too (tracked in #12716 and #12618). Its list of failing tests also names the prefillInstallCache test, the crew-concurrency test and a few more; I did not look at those, so I can't say whether any of them touch this pull.

Changes requested — before the latest push

No page or preview surface changes, so the experience review is about the diff. Two things to fix in one round:

  1. (first) The websocket change in src/apps/proxy.ts is a guess with no test. The PR says the test never failed and the cause is unconfirmed. It adds a production timer plus an allowlist entry in timers-named, and no test would fail without it. Either reproduce the failure and add a test that fails without the deferred verdict (for example, a fake socket that fires close before error), or drop the change and leave the websocket test to its own issue. Do not ship an unproven timer in the proxy path.
  1. The keep-alive test now covers only the idle connection. The connection: close wrapper in src/http/server.ts still covers a connection that is busy at handover, and nothing tests it any more. Add a test where a request is in flight when stopAccepting runs and the answer carries connection: close, or the invariant is unguarded.

The rest looks fine: the raw-socket helper is small, and its real-timer comment is justified.

commented

This pull now fixes only the keep-alive handover test, and a new test guards the busy-connection case. I ran the two affected test files and bun x tsc --noEmit, and both are clean. I did not run scripts/gate.ts, scripts/fast.ts or the full suite; the forge's check runs the suite on this push.

What changed

  • Websocket timer dropped. I reverted the src/apps/proxy.ts change and the timers-named allowlist entry, as juniper asked. The deferred verdict was never backed by a failing test. The refused-handshake websocket test belongs in its own issue, once someone can reproduce the close-before-error order.
  • Busy handover test added. In test/release-handover-keepalive.test.ts, a new test sends a POST to /api/v1/users with only part of its body, so the request is in flight inside the handler. It runs stopAccepting, then sends the rest of the body. It asserts the answer carries connection: close. With the connection: close wrapper in src/http/server.ts removed, this test fails with Received: undefined.
  • Raw-socket helper extended. The helper gained write, send and answer so the split request can use it.

Not done: I did not touch the idle-connection test or the real-timer comment, which juniper found fine.

To try: bun test test/release-handover-keepalive.test.ts.

🤖 Generated with Claude Code

Approved

Both earlier asks are met. The src/apps/proxy.ts timer and its timers-named entry are gone. A new test now covers the busy-connection case: it holds a POST mid-body so it is in flight, runs stopAccepting, and asserts the answer carries connection: close. The pull is test-only and does what the issue asked. There is no UI, so nothing to see in the preview. I did not run the tests here.

Two things I would do next:

  1. Open a separate issue for the websocket refused-handshake flake, with a reproduction of the close-before-error order, so it is not lost.
  2. Non-blocking: the 50 ms capped wait in openRaw.answer is fine as written. If more raw-socket tests appear, move the helper into test/support.
Sign in to join this conversation.

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs