Skip to content

relay restart crashes whole Chromium tabs that hold a WebTransport session; the published @moq/net only logs the relay's GOAWAY #4785

Closed

Description

@aaisino

Summary

When we restart moq-relay (systemctl restart, i.e. SIGTERM), Chromium tabs connected over WebTransport crash outright (renderer SIGSEGV), not just disconnect. The same tabs forced onto WebSocket almost never crash. By its source the relay answers SIGTERM with a GOAWAY to every session and force-closes what is left after drain_timeout (10 s by default), but the published @moq/net 0.4.1 only logs a GOAWAY, so the sessions are closed under the tabs. The client migration is on release (b4234f8, #4132) but in no npm package, and the docs do not say that restarting a relay under live WebTransport clients can take their tabs down. We suspect a Chromium bug in the teardown of a server-closed WebTransport session; we have not reduced it to a Chromium test case, and we have not tested a client that migrates.

Versions

Relay 0.15.8; clients @moq/net 0.3.5 and 0.4.1; Edge 154 (headed) and headless Chromium 151, QUIC over a real network. Not run against relay 0.17.0 or a client with the migration.

Steps to reproduce

  1. Two browser tabs, each publishing and subscribing, on a relay over WebTransport.
  2. systemctl restart moq-relay, three times; count the tabs that crash.
  3. Repeat with WebSocket forced.

Expected

Restarting a relay does not take tabs down: clients act on the GOAWAY and migrate or close the session cleanly before the relay exits, or the docs say what to expect and how to restart safely.

Actual

  • Edge 154 (headed): 4 of 6 tabs crashed, with each of the two client lines.
  • Headless Chromium 151: 1 of 6 (@moq/net 0.3.5) and 2 of 6 (0.4.1).
  • WebSocket: 0-1 of 12 tabs.
  • A tab that survives reconnects after 0.6-1.9 s; video and audio are back after 2.3-3.3 s.
    We did not capture at which moment of the drain the tabs die.

Where in the code

  • Published @moq/net 0.4.1: js/net/src/lite/connection.ts:253-255 logs received goaway: and does nothing else; IETF: js/net/src/ietf/adapter.ts:271-272 and js/net/src/ietf/connection.ts:329 only console.warn.
  • Migration on release (moq-relay-v0.17.0): js/net/src/connection/goaway.ts, js/net/src/connection/reload.ts, js/net/src/connection/pool.ts; doc/lib/js/net.md:51 describes it (dial the replacement at once, keep the old session until goaway.handover, 10 s by default).
  • Relay side: rs/moq-relay/src/config.rs:92-104 (drain_timeout, "The first signal sends every session a GOAWAY and waits"), doc/bin/relay/config.md:242-263 ("Only moq-lite-04+ and moq-transport clients act on a GOAWAY; older ones are closed when the window ends").

Workaround we use

We restart the relay only when nobody is connected.

Related

Activity

  1. kixelated commented on Oct 4, 2026

    @kixelated
    Collaborator

    Thanks for the measurements. The client side has shipped: GOAWAY migration (#4132) is in @moq/net 0.4.2, published today, so a client now moves off a draining relay instead of being force-closed at drain_timeout. A renderer SIGSEGV can't be caused legitimately by page JavaScript, so if tabs still crash with 0.4.2 against relay 0.17, that is a Chromium bug in WebTransport teardown, and a crbug with a minimal repro is the right home for it. Closing; please reopen with the results if a migrating client still crashes.

    (written by Claude Opus 5.5)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions