auto-rebase dies in job setup when data.forgejo.org is unreachable - 16 consecutive ticks, 7.5 h with no rebase or arming #2366
Labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
jaylouisw/packetforge#2366
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
What happened
On 2026-11-20 the runner lost egress to
data.forgejo.org. Everyauto-rebase.ymltick in that window failed before its first step:Each one:
Set up job2-3 s, then every real stepcancelled, jobfailure. The whole log is 12 lines (run 8490, jobs/0/attempt/1/logs):Root cause
The runner pre-clones every
uses:action of a job at dispatch, before step one.auto-rebase.ymlis the only workflow left in this repository with auses:(actions/checkout@v4, line ~305), so it is the only one that cannot start when the internet is down.ci.ymlremoved that dependency deliberately (its checkout isgit init+git fetch --depth=1, see the note at the top ofci.yml), which is why the test legs, smoke, docker and arm kept running through the same outage - and why the docker-CLI fallout of the same outage was already fixed inci.yml(task 22378, run 8465).Cost: 7 h 20 m in which no stale pull request was rebased and the scheduled arming backstop did not run, plus 16 red runs that read like code failures.
Why the workflow cannot fix it itself
The job has to keep a checkout step:
test/ci-disk-hygiene.test.js(case #1508, lines ~186-190) asserts that therebasejob contains a step matchinguses:\s*actions/checkout@, positioned after thedf -Pkroom guard and gated on itsroomoutput. Anyuses:- even one behindif: false- is pre-fetched at dispatch, so the setup failure is unavoidable while that line exists.Proposed patch (needs the test/ owner)
auto-rebase.yml: replace theuses: actions/checkout@v4step with the bloblessgit init+git fetch --depth=1 <sha>the rest of the repository already uses (pr-rebase.yml,close-merged-issues.yml). Same commit, same tree, samegit clean -qfdx, still gated on the room guard; removes ~3.8 s of action clone + ~9.2 s of ref fetch per tick as well.test/ci-disk-hygiene.test.js: make the invariant "the room guard precedes whatever checks out", not "the job uses the checkout action" - accept a step whose text hasuses: actions/checkout@orgit+fetch --depth=1, keep theguardAt < checkoutAtandgated()assertions unchanged.No coverage changes: the tick still rebases the same branches onto the same base with real git, and still arms through the same API calls.
Not affected
pr-rebase.yml(event-triggered rebase) has nouses:and kept working;ci.yml's per-PR arm job kept arming after each matrix.Fixed in #2374 (branch
fix/1121-0532). Both halves of the proposal landed: theuses: actions/checkout@v4step is nowgit init+ one blobless fetch ofmain, gated exactly as before, with credentials added tooriginonly for the push; andtest/ci-disk-hygiene.test.jsnow reads the invariant as "the room guard precedes whatever materialises the tree", plus a new case that fails onmainand asserts therebasejob runs nouses:at all, so this cannot come back.npm test: 435 suites, 14624 assertions, 0 failed. The loop was replayed against a local bare repo from an object-less clone - fetch,checkout -B,merge-base --is-ancestor,rebase,diff --quietand--force-with-leasepush all succeed, blobs fetched on demand.