Repository navigation
fix(gateway): keep a selected peer fence through deploy recovery - #5634
Conversation
restore-disabled and verify assumed the fence is never selected, so the first revision selecting it would have had its fence removed or its deploy failed. Both now follow the checked-out source, and an unreadable selection fails without a write. The fence stays unselected, so nothing changes on live routes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@coderabbitai review |
Action not completed
|
|
No actionable comments were generated in the recent review.
|
A single key in one kustomization missed a selection made from a parent overlay, which would have removed a selected fence. An independent review reproduced that with real renders. The render also has to contain the prepared policies, so reading the wrong source fails instead of reading as unselected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@coderabbitai review |
Action performed
|
Readiness record at
|
Correction to the readiness record above. It said the unselected path would first run for real in this PR's merge-group deploy. That was wrong: in run 41238112849 the |
The live check owed above is done. The merge-group deploy of #5635 (run 41239030101), which is one commit ahead of this merge, ran the new helper against production and logged |
Why
Email recovery needs the physical-peer fence selected before the proxy transport can be turned on, but every deploy currently removes that fence and then fails unless it is absent. Selecting it today would undo itself or evict its own pull request.
What
The deploy's restore and verify steps now follow what the checked-out source selects: unselected keeps today's behaviour, selected writes nothing and requires the reviewed fence to be live. The fence stays unselected here, so public routes do not change; selecting it is the next change.
Part of #5448