Skip to content

recover/re-copy: finalize not idempotent for already-published datasets (Dataset is already published) #985

Closed

Description

@neuromechanist

Found during the epic #967 recovery canary (on002814, prod, 0.9.1) — before running the 45-batch.

Symptom: nemar admin recover on002814 --execute dispatched onboard-openneuro.yml, which ran cleanly through parse-ids -> prepare -> copy (all 8 shards) -> finalize, where it FAILED:

✔ Verified 44895 objects present
✖ Failed to request publication: Dataset is already published
##[error]Process completed with exit code 1

The re-copy itself WORKED: after the run, nemar admin import verify on002814 reports verified complete (44909/44909 objects present) and the dataset serves HTTP 200 at data.nemar.org. The 0-byte data is fixed.

Root cause: finalizeImport (the onboard-openneuro finalize step) unconditionally calls requestPublication. For a RE-copy of an already-published dataset (every one of the 45 #967 recover targets), that fails with 'Dataset is already published' and process.exit(1). Phase 2 made prepare idempotent for re-dispatch (Step 4b), but the finalize publication path was never made idempotent.

Impact:

  1. The workflow run reports failure even though the data was re-copied (misleading; noisy across a 45-batch).
  2. Any finalize step AFTER requestPublication is skipped on the exit(1) — so datasets.data_complete stays stale/unaudited, and any manifest/archive/records regeneration that lives past that point does NOT run. So the S3 data is fixed but downstream artifacts + the catalog completeness signal are not refreshed. (on002814's data_complete is still unaudited after the run.)
  3. Affects ALL 45 recover targets (all already-published), so the batch would produce 45 'failed' runs with fixed data but stale bookkeeping.

Fix: in the recover/re-copy finalize path, detect an already-published dataset and SKIP (or treat as success) the publication request, then continue to the per-key verify + completeness/manifest updates. A re-copy of a published dataset should be a data-repair, not a re-publish. Add a real test for the already-published re-copy path.

Interim: after a re-copy lands the data, data-integrity-sweep --older-than 0 sets data_complete correctly (independent of finalize), so the catalog signal can be repaired separately.

Blocks the clean batch recovery of the 45. Part of the #967 follow-up set (with #984).

Activity

  1. added 2 commits that reference this issue on Jul 22, 2026
  2. neuromechanist commented on Aug 3, 2026

    @neuromechanist
    ContributorAuthor

    Fixed and shipped: PR #986 (de9ec36, "fix: skip publish request/approve for already-published re-copy"), released in v0.9.2 on 2026-07-22 and on main since.

    finalizeImport used to request publication unconditionally, so a #967 recover re-copy of an already-published dataset failed with "Dataset is already published" after it had verified the data was present. src/lib/import-openneuro.ts now decides via decidePublicationRequestAction(visibility) — public returns skip-already-published, anything else returns request — keyed on the dataset's live visibility rather than on assuming a first-time import. Covered in test/import-openneuro.test.ts.

    Closing. (PRs merged to dev never auto-close issues here — the default branch is main.)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions