deploy.yml's concurrency group cannot collapse a release push: annotated tags give every ref a different github.sha, so three identical builds run in parallel
#1338
SAME ASSUMPTION AS archie-core-i5nb, QUIETER CONSEQUENCE. i5nb was a ref's object id being consumed as a commit and mislabelling an artifact; this is the same assumption costing duplicate work instead. Found while fixing i5nb in the same file; deliberately NOT fixed there.
THE INTENT, from the file's own comment above the group:
'# Keyed on the commit, not github.ref: git push --follow-tags pushes main plus both archied/v* and archie/v* tags in one go, and all three refs point at the same commit -- on.push matches branches and tags both, so a single release push otherwise fires three redundant builds of identical content that a ref-keyed group can't see as the same work.'
WHAT ACTUALLY HAPPENS
concurrency:
group: deploy-${{ github.sha }}
github.sha is the ref's OBJECT id, not a commit. Every release tag in this repo is ANNOTATED — measured: 80 of 80 archied/v* and archie/v* tags are tag objects, zero lightweight — so on a release push the three refs resolve to three DIFFERENT ids: main gives the commit, archied/vX gives one tag object, archie/vX gives another. Three groups, so the three redundant builds the comment exists to prevent run in parallel. Before a release, when only main is involved, the group is unique per commit and works exactly as intended, which is why this has not been noticed.
Evidence: archied/v1.36.0 -> tag object 6720bc4249effccc46ea17d3169ea089856d0b1f, dereferences to commit c25f8d3956.
WHY IT IS NOT A PATCH — TWO ALTERNATIVES THAT DO NOT WORK, recorded so they are not re-derived
Resolve the commit in a step and use that as the group. Impossible: 'concurrency' is evaluated BEFORE any step runs, so no step can compute the key. This is the fact that makes it a design question rather than an edit.
Key on the ref instead. Worse than the bug: with 'cancel-in-progress: true', the three release builds would cancel each other, so a release push would lose work rather than duplicate it.
There is no GitHub expression that dereferences a tag object, so this needs a deliberate design (a fixed group plus an explicit guard, or accepting ref-keyed groups with cancel-in-progress off, or normalising the push — the choice is the maintainer's).
WHAT TO PRESERVE WHEN SOLVING IT: the property the comment asks for — main plus both release tags at one commit must land in ONE group, so a release is built once, and a newer push must still supersede an older one.
RELATED: archie-core-i5nb, fixed in e2df486d for the image tags (each build now publishes sha- and the commit is obtained by dereferencing the ref). That commit also added the 80-of-80 measurement next to the dereference step.
Close reason
Invalid premise, verified 2026-10-04: annotated v1.46.0 peels to b7d7a9ea; deploy runs 36958439371 (tag, success) and 36958438706 (main, cancelled) share that head SHA. v1.45.0 and v1.44.x similarly share commit SHAs and cancel one run. Current deploy concurrency already collapses these runs. GitHub documents push GITHUB_SHA as the tip commit, not the annotated tag object. The legacy three-tag release premise is also superseded by one release stream.
Imported from beads archie-core-gixd · GitHub #1094 · type bug · priority P2 · created 2026-09-22 · closed 2026-10-04
SAME ASSUMPTION AS archie-core-i5nb, QUIETER CONSEQUENCE. i5nb was a ref's object id being consumed as a commit and mislabelling an artifact; this is the same assumption costing duplicate work instead. Found while fixing i5nb in the same file; deliberately NOT fixed there.
THE INTENT, from the file's own comment above the group:
'# Keyed on the commit, not github.ref: git push --follow-tags pushes main plus both archied/v* and archie/v* tags in one go, and all three refs point at the same commit -- on.push matches branches and tags both, so a single release push otherwise fires three redundant builds of identical content that a ref-keyed group can't see as the same work.'
WHAT ACTUALLY HAPPENS
concurrency:
group: deploy-${{ github.sha }}
github.sha is the ref's OBJECT id, not a commit. Every release tag in this repo is ANNOTATED — measured: 80 of 80 archied/v* and archie/v* tags are tag objects, zero lightweight — so on a release push the three refs resolve to three DIFFERENT ids: main gives the commit, archied/vX gives one tag object, archie/vX gives another. Three groups, so the three redundant builds the comment exists to prevent run in parallel. Before a release, when only main is involved, the group is unique per commit and works exactly as intended, which is why this has not been noticed.
Evidence: archied/v1.36.0 -> tag object 6720bc4249effccc46ea17d3169ea089856d0b1f, dereferences to commit c25f8d39566b78507bcd85b22d0b9e00f3870a75.
WHY IT IS NOT A PATCH — TWO ALTERNATIVES THAT DO NOT WORK, recorded so they are not re-derived
1. Resolve the commit in a step and use that as the group. Impossible: 'concurrency' is evaluated BEFORE any step runs, so no step can compute the key. This is the fact that makes it a design question rather than an edit.
2. Key on the ref instead. Worse than the bug: with 'cancel-in-progress: true', the three release builds would cancel each other, so a release push would lose work rather than duplicate it.
There is no GitHub expression that dereferences a tag object, so this needs a deliberate design (a fixed group plus an explicit guard, or accepting ref-keyed groups with cancel-in-progress off, or normalising the push — the choice is the maintainer's).
WHAT TO PRESERVE WHEN SOLVING IT: the property the comment asks for — main plus both release tags at one commit must land in ONE group, so a release is built once, and a newer push must still supersede an older one.
RELATED: archie-core-i5nb, fixed in e2df486d for the image tags (each build now publishes sha-<commit> and the commit is obtained by dereferencing the ref). That commit also added the 80-of-80 measurement next to the dereference step.
### Close reason
Invalid premise, verified 2026-10-04: annotated v1.46.0 peels to b7d7a9ea; deploy runs 36958439371 (tag, success) and 36958438706 (main, cancelled) share that head SHA. v1.45.0 and v1.44.x similarly share commit SHAs and cancel one run. Current deploy concurrency already collapses these runs. GitHub documents push GITHUB_SHA as the tip commit, not the annotated tag object. The legacy three-tag release premise is also superseded by one release stream.
---
_Imported from beads `archie-core-gixd` · GitHub #1094 · type bug · priority P2 · created 2026-09-22 · closed 2026-10-04_
sam
added the mediumbug labels 2026-10-27 13:39:59 +00:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
SAME ASSUMPTION AS archie-core-i5nb, QUIETER CONSEQUENCE. i5nb was a ref's object id being consumed as a commit and mislabelling an artifact; this is the same assumption costing duplicate work instead. Found while fixing i5nb in the same file; deliberately NOT fixed there.
THE INTENT, from the file's own comment above the group:
'# Keyed on the commit, not github.ref: git push --follow-tags pushes main plus both archied/v* and archie/v* tags in one go, and all three refs point at the same commit -- on.push matches branches and tags both, so a single release push otherwise fires three redundant builds of identical content that a ref-keyed group can't see as the same work.'
WHAT ACTUALLY HAPPENS
concurrency:
group: deploy-${{ github.sha }}
github.sha is the ref's OBJECT id, not a commit. Every release tag in this repo is ANNOTATED — measured: 80 of 80 archied/v* and archie/v* tags are tag objects, zero lightweight — so on a release push the three refs resolve to three DIFFERENT ids: main gives the commit, archied/vX gives one tag object, archie/vX gives another. Three groups, so the three redundant builds the comment exists to prevent run in parallel. Before a release, when only main is involved, the group is unique per commit and works exactly as intended, which is why this has not been noticed.
Evidence: archied/v1.36.0 -> tag object 6720bc4249effccc46ea17d3169ea089856d0b1f, dereferences to commit
c25f8d3956.WHY IT IS NOT A PATCH — TWO ALTERNATIVES THAT DO NOT WORK, recorded so they are not re-derived
There is no GitHub expression that dereferences a tag object, so this needs a deliberate design (a fixed group plus an explicit guard, or accepting ref-keyed groups with cancel-in-progress off, or normalising the push — the choice is the maintainer's).
WHAT TO PRESERVE WHEN SOLVING IT: the property the comment asks for — main plus both release tags at one commit must land in ONE group, so a release is built once, and a newer push must still supersede an older one.
RELATED: archie-core-i5nb, fixed in
e2df486dfor the image tags (each build now publishes sha- and the commit is obtained by dereferencing the ref). That commit also added the 80-of-80 measurement next to the dereference step.Close reason
Invalid premise, verified 2026-10-04: annotated v1.46.0 peels to b7d7a9ea; deploy runs 36958439371 (tag, success) and 36958438706 (main, cancelled) share that head SHA. v1.45.0 and v1.44.x similarly share commit SHAs and cancel one run. Current deploy concurrency already collapses these runs. GitHub documents push GITHUB_SHA as the tip commit, not the annotated tag object. The legacy three-tag release premise is also superseded by one release stream.
Imported from beads
archie-core-gixd· GitHub #1094 · type bug · priority P2 · created 2026-09-22 · closed 2026-10-04