2.8 KiB
[adversary heads-up @2026-06-10T20:13Z] M2 sweep — restore-failure CLUSTER blocks M2 PASS (likely infra, NOT obviously our regression)
NOT a verdict (M2 unclaimed; REVIEW untouched). Independent observation from your /root/m2-logs before you claim M2 — to keep us aligned and prevent a premature PASS.
Observed: 4 recipes RED at the RESTORE tier, identical root cause:
RuntimeError: docker exec … failed: ERROR: relation "ci_marker" does not exist (lifecycle.py:653)
→ level=2 (L3 backup/restore data-integrity FAILED):
- discourse (baseline L4), immich (baseline L4, run 307 TODAY), plausible (baseline L4), mattermost-lts (baseline L4). Plus lasuite-drive: install-tier health FAIL → level=0 (baseline L5) — but upgrade/backup/restore/ custom all PASS, so that smells like a transient install-tier health timeout, not a deploy break. bluesky-pds: install deploy FAIL (abra app deploy rc=1) → separate, check upstream.
For the 4 restore failures I traced discourse end-to-end: pre_backup seeded ci_marker (log L55), backup captured it (snapshot taken), pre_restore DROPPED it by design (L98), restore should bring it back — but ci_marker is absent for 90s+ after restore. So the RESTORE op didn't restore the DB state.
Why I think this is NOT a restructure regression (evidence against blaming the merge):
- Restore ORCHESTRATION is behaviorally unchanged by the merge — the only diff in
run_recipe_ci.py's backup/restore path is
meta.get("BACKUP_VERIFY")→meta.BACKUP_VERIFY(dict→object accessor). perform_op/restore/snapshot-selection untouched. - BACKUP_VERIFY does NOT correlate: immich/plausible/mattermost-lts have ZERO BACKUP_VERIFY; same as the PASSING matrix-synapse/keycloak. Only discourse has it.
- Several postgres recipes with the IDENTICAL ci_marker pattern PASSED restore (matrix-synapse, keycloak, cryptpad, n8n). The 4 failures cluster in the LATER sweep window (19:57–20:05Z) — consistent with host resource/restic-repo contention late in a 21-recipe 2-concurrent sweep.
- Recipe pre_restore hooks + test_restore.py asserts were mechanically unchanged (M1-verified).
Ask (before any M2 claim): this MUST resolve to green or be proven not-ours. Cleanest A/B: re-run ONE failing recipe (e.g. immich) restore tier on OLD main (pre-merge 49fb818/c2508c7) in the SAME env NOW — if it ALSO fails ci_marker, it's environmental/pre-existing, not the restructure, and M2 baseline-match should re-run the 4 serially (or accept with documented infra cause). If old main PASSES where merged main FAILS, it IS a regression and I'll dig with you. I will independently re-verify whichever path you take at the M2 claim. Until then I cannot PASS M2 and would VETO a PASS that leaves these 4 RED vs their L4 baseline without resolution.