|
| 1 | +--- |
| 2 | +"@objectstack/objectql": minor |
| 3 | +--- |
| 4 | + |
| 5 | +feat(objectql): report dangling lookup references left behind by `isSystem` writes (#4551) |
| 6 | + |
| 7 | +#4441 made the write path refuse a `lookup` id that exists in no row of the |
| 8 | +object it references, but deliberately exempted `isSystem` writes: seed replay, |
| 9 | +package install and boot-time provisioning legitimately write rows in an order |
| 10 | +that is self-consistent only once the batch completes, and failing them closed |
| 11 | +would turn an ordering detail into a boot failure. The exemption is correct, and |
| 12 | +it left a real residual — **the platform itself could still store a reference |
| 13 | +pointing at nothing, and nothing said so.** PR #4511 recorded that residual |
| 14 | +rather than silently accepting it; this closes it, by reporting. |
| 15 | + |
| 16 | +`ObjectQL.inspectDanglingReferences()` walks the non-`readonly` reference fields |
| 17 | +of every registered object and names each stored id that resolves to no row: |
| 18 | +object, record id, field, target object, the id itself, and the holding row's |
| 19 | +tenant. It **reports and never rewrites** — the rows genuinely exist and were |
| 20 | +genuinely written, so nulling a link would make the stored data disagree with |
| 21 | +what happened, and whether the repair is to create the missing target, re-run |
| 22 | +the seed or delete the holder is a judgement the platform cannot make. |
| 23 | + |
| 24 | +Three properties are worth knowing before reading a report: |
| 25 | + |
| 26 | +- **Unknown is not healthy.** The probe is the same one the write path enforces |
| 27 | + with. It answers "missing" only when it RAN and found nothing; an unregistered |
| 28 | + target, an absent driver or a probe that throws is counted into |
| 29 | + `undetermined`, and an unreadable object is named in `skipped`. A datasource |
| 30 | + outage can therefore never read as a clean bill of health. |
| 31 | +- **`readonly` references are skipped**, as on the write path: the value there |
| 32 | + was minted by the platform, and at least one is a sentinel by design |
| 33 | + (`sys_metadata_history.recorded_by` stores `actor ?? 'system'` in a |
| 34 | + `lookup('sys_user')`). |
| 35 | +- **Bounded, RBAC-first.** The RBAC link tables are visited first — a dangling |
| 36 | + row there is a security-surface record that resolves to nothing, and the |
| 37 | + audience-anchor gate has to resolve that very target to evaluate the grant. |
| 38 | + Per-object and total row caps bound one pass; hitting one sets `truncated`, so |
| 39 | + a report that stopped early cannot read as "everything was checked". |
| 40 | + |
| 41 | +It rides the existing ADR-0057 lifecycle clock (hourly, first run delayed past |
| 42 | +boot — exactly when seed-written references become checkable) rather than arming |
| 43 | +a second one, so the finding surfaces without an operator knowing to go looking. |
| 44 | +It runs *after* the sweep and is isolated from it in both directions. |
| 45 | +`LifecycleService.sweep()` is unchanged: tooling that calls it directly still |
| 46 | +gets policy enforcement and nothing else. |
| 47 | + |
| 48 | +New exports from `@objectstack/objectql`: `DanglingReference`, |
| 49 | +`DanglingReferenceReport`, `REFERENCE_SCAN_PRIORITY_OBJECTS`, |
| 50 | +`DANGLING_SCAN_ROWS_PER_OBJECT`, `DANGLING_SCAN_MAX_ROWS`. Nothing was removed |
| 51 | +or renamed, and #4441's enforcement is untouched. |
0 commit comments