Timeout Missing On One Step

What goes wrong

A timeout bounds the main call but is not passed to a prerequisite or follow-up call made in the same operation — fetching a token, opening the connection, an external check — so that step can hang past the promised ceiling and the worker or connection slot stays occupied indefinitely.

Example

The HTTP request has a 5-second timeout, but the token refresh before it has none; the auth server stalls and the request hangs for minutes.

Not to be confused with

The inner call gets no deadline at all; not a budget too small for its known phases (Outer deadline omits inner work), and not work left running after its deadline fires (Abandoning Deadline).

Discovered by

@sks24git GitHub

Field notes · 5

@sks24git2026-11-04beat 3 review rounds

A network call that ran only inside a step with a deadline was reused in a step without one, and it has no timeout of its own. When the remote service hangs, the second step hangs with it.

Give every outbound call its own timeout. A caller's deadline does not travel when the call is reused.
Check what a reused helper relied on. Its old callers' guards stay behind.

@sks24git2026-11-03beat 3 review rounds

An exit-time usage write went through a connection pool, where waiting for a free slot is not covered by the query timeout. A busy pool could hold every exiting process far past the limit the docs promised.

Use a direct, short connection at exit. Pool slot waits escape statement timeouts.
Count every statement in the exit budget. Each one gets its own timeout, and they add up.

@sks24git2026-11-03beat 2 review rounds

An exit hook that reports usage gained a database write with a connect timeout but no query timeout. A slow or locked database now holds every process at exit, and a kill during the write prints the report twice.

Bound every step of exit-time I/O. Connect, lock and query each need their own limit.
Mark the report done before slow work. A signal mid-write then does not repeat it.

@sks24git2026-11-03beat 1 review round

A startup cleanup takes a row lock with no time limit, while every other startup step that locks has one. If any live transaction holds that row, the server never finishes starting.

Bound every lock wait on the startup path. A boot that queues behind a live transaction is an outage.

@sks24git2026-11-01beat 3 review rounds

Required merge checks moved to a small runner pool, and the job timeout only counts run time, not queue time. A burst of pushes or a dead pool leaves every merge waiting on checks that never start.

Bound queue time, not just run time. A job timeout never fires while the job waits for a runner.
Alert on an empty pool. A required check on a dead pool blocks merges in silence.