Files
flatpark/scripts/drop-dangling-refs.sh
T
KomhandClaude Opus 5 bb35120069 fix(ci): reclaim the ref files a de-list leaves behind in R2
publish uploads refs with `rclone copy`, which never deletes, so a ref that
delist-prune removed locally survived in R2. Once the weekly prune reclaimed
that commit's objects the leftover file became a dangling pointer, and the next
publish's R2 reconcile synced it straight back into the local repo — where the
summary regeneration flatpak build-export does on export hard-fails:

    Exporting app.markra.Markra to repo
    error: No such metadata object 84604b86...aac26297.commit

That commit was io.github._66HEX.Frame's, de-listed on 2026-07-27 and pruned on
08-01; app.legcord.Legcord left an identical landmine. Nothing about the app
being built was wrong, and no rerun could pass: every publish after a de-list +
prune fails the same way.

- prune-and-reclaim.sh now records the ref files it delists, and sync-r2.sh
  deletes those keys from R2 after the new summary is live (refs before the
  objects they name) — same targeted-delete discipline as the object reclaim,
  still never a blind mirror.
- prune-and-reclaim.sh also delists runtime/<id>.{Debug,Locale,Sources}, not
  just app/<id>: a de-listed app's .Debug ref used to stay live forever and pin
  its objects against reclaim (app.legcord.Legcord.Debug is still in the summary
  today). Genuine runtimes are left alone.
- drop-dangling-refs.sh sweeps after the R2 reconcile so a publish self-heals on
  stale state instead of wedging. It queues the R2 delete only for refs whose
  app left the registry; a ref that is still listed is dropped locally only, so
  a half-finished object sync can never cost a published pointer.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
2026-08-02 21:19:03 +08:00

60 lines
2.9 KiB
Bash
Executable File

#!/usr/bin/env bash
# Drop refs whose commit object is not in the repo, and record the ref files so
# the caller can delete the same paths from R2 (see sync-r2.sh RECLAIM_REFS).
#
# Why this exists: sync-r2.sh uploads refs with `rclone copy`, which never
# deletes, so a ref that delist-prune removed locally survives in R2. Once
# prune reclaims that commit's objects, the leftover ref file is a dangling
# pointer — and publish's R2 reconcile syncs it straight back into the local
# repo, where the next summary regeneration (flatpak build-export does one on
# export) hard-fails with "No such metadata object <commit>.commit". That wedges
# every publish, not just the one after the de-list.
#
# Deleting here stays within the "targeted delete, never a blind mirror" rule:
# only refs whose commit object is provably absent are touched, and the local
# object store is an exact copy of R2's at this point in the run.
ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
. "$ROOT/scripts/lib/common.sh"
load_config "$ROOT"
need ostree
repo="${1:-$REPO_DIR}"
[ -d "$repo/objects" ] || die "not an ostree repo: $repo"
# Object stores do not preserve empty directories; R2 omits refs/remotes/ when
# it is empty and `ostree refs` fails without it.
mkdir -p "$repo/refs/remotes" "$OUT_DIR"
# Truncated, not appended: this sweep runs first and owns the list;
# prune-and-reclaim.sh appends the refs it delists to the same file.
reclaim_refs="${RECLAIM_REFS:-$OUT_DIR/reclaim-refs.txt}"
: > "$reclaim_refs"
dropped=0
while IFS= read -r ref; do
[ -n "$ref" ] || continue
case "$ref" in *:*) continue ;; esac # remote ref, not ours to touch
commit="$(ostree --repo="$repo" rev-parse "$ref" 2>/dev/null || true)"
[ -n "$commit" ] \
&& [ -f "$repo/objects/${commit:0:2}/${commit:2}.commit" ] \
&& continue
warn "dangling ref $ref -> ${commit:-<unreadable>}: commit object missing, dropping"
# Queue the R2 delete only for a ref whose app left the registry — that is
# the one case where the leftover file is provably garbage. A live app with
# a dangling ref (a half-finished object sync, say) is only dropped locally:
# its next build re-exports the ref and overwrites R2's copy, so a bad sync
# can never cost us a published pointer.
if ref_is_delisted "$ref"; then
# Record the file that actually holds the ref (heads or mirrors), so the
# R2 delete targets the same key the reconcile would pull back down.
for base in refs/heads refs/mirrors; do
[ -f "$repo/$base/$ref" ] && printf '%s\n' "$base/$ref" >> "$reclaim_refs"
done
else
warn "keeping R2's copy of $ref: still in the registry, a rebuild will replace it"
fi
ostree --repo="$repo" refs --delete "$ref"
dropped=$((dropped + 1))
done < <(ostree --repo="$repo" refs)
log "dangling refs dropped: $dropped -> $reclaim_refs"
printf '%s\n' "$reclaim_refs"