Restore preserved states when a deletion is cancelled
What does this MR do and why?
Problem: When you cancel a group's deletion, all subgroups become "active" — even ones that were archived before the deletion was scheduled.
Solution: This MR remembers each subgroup's original state and restores it when deletion is cancelled.
State Transition Overview
stateDiagram-v2
direction LR
ancestor_inherited --> archived: archive
ancestor_inherited --> deletion_scheduled: schedule deletion
archived --> deletion_scheduled: schedule deletion
archived --> ancestor_inherited: unarchive
deletion_scheduled --> ancestor_inherited: cancel (restore)
deletion_scheduled --> archived: cancel (restore to archived)Example Scenario
Before this MR:
Parent Group (deletion_scheduled → ancestor_inherited)
├── Subgroup A (was archived before deletion) → ❌ becomes ancestor_inherited
└── Subgroup B (was active before deletion) → ancestor_inheritedAfter this MR:
Parent Group (deletion_scheduled → ancestor_inherited)
├── Subgroup A (was archived before deletion) → ✅ restored to archived
└── Subgroup B (was active before deletion) → ancestor_inheritedTechnical Details
- Each descendant goes back to its preserved
schedule_deletionstate, but never below the parent's new state. A descendant with nothing preserved takes the parent's state. The preserved state is then removed. - Cancelling an archived group's deletion (
deletion_scheduled→archived) is now a reversal, so itsdeletion_scheduleddescendants are restored too, as ADR 003 describes ("Allowed as restore to archived state").
Nothing creates propagation records yet (#599305 (closed)), so this has no user-facing effect until that lands.
Database queries
Restore mode runs these per batch of up to 200 namespace IDs. Timings are from a gitlab-production-main Database Lab thin clone, using a batch of 200 IDs spread across the ID range (133 matched rows). The UPDATE was measured as a single statement covering the whole batch, the worst case. The existing query that removes preserved states is not re-measured.
Lock the batch's rows. Restore now matches only rows still in deletion_scheduled, so a re-run after a partial failure skips descendants it already restored. Measured with 200 randomly chosen deletion_scheduled IDs, so every row matches and is locked (the worst case):
SELECT "namespaces"."id", "namespaces"."state"
FROM "namespaces"
WHERE "namespaces"."state" = 2
AND "namespaces"."id" IN (...)
ORDER BY "namespaces"."id" ASC
LIMIT 200
FOR NO KEY UPDATELimit (cost=0.57..555.70 rows=1 width=12) (actual time=0.061..1.477 rows=200 loops=1)
Buffers: shared hit=1395
-> LockRows (cost=0.57..555.70 rows=1 width=12) (actual time=0.060..1.460 rows=200 loops=1)
Buffers: shared hit=1395
-> Index Scan using namespaces_pkey on public.namespaces (cost=0.57..555.69 rows=1 width=12) (actual time=0.021..1.158 rows=200 loops=1)
Index Cond: (namespaces.id = ANY ('{...}'::integer[]))
Filter: (namespaces.state = 2)
Buffers: shared hit=1000- Cold: 547 ms (534 buffers read, ~4.2 MiB; 224 dirtied by the row locks). Warm: 4.1 ms (1,395 hits).
- Plan: https://console.postgres.ai/gitlab/projects/gitlab-production-main/sessions/69019/commands/173690
Read each namespace's preserved state:
SELECT "namespace_details"."namespace_id",
state_metadata -> 'preserved_states' ->> 'schedule_deletion'
FROM "namespace_details"
WHERE "namespace_details"."namespace_id" IN (...)Index Scan using namespace_details_pkey on public.namespace_details (cost=0.57..676.43 rows=200 width=40) (actual time=0.053..1.607 rows=133 loops=1)
Index Cond: (namespace_id = ANY ('{...}'::bigint[]))
Buffers: shared hit=933- Cold: 584 ms (524 buffers read, ~4.1 MiB). Warm: 2.8 ms (933 hits)
- Plan: https://console.postgres.ai/gitlab/projects/gitlab-production-main/sessions/66985/commands/171794
Move rows to their restored state, one statement per distinct state (at most two: ancestor_inherited and archived):
UPDATE "namespaces" SET "state" = 1 WHERE "namespaces"."id" IN (...)Update on public.namespaces (cost=0.57..556.92 rows=0 width=0) (actual time=776.890..776.891 rows=0 loops=1)
Buffers: shared hit=10059 read=589 dirtied=1082 written=10
-> Index Scan using namespaces_pkey on public.namespaces (cost=0.57..556.92 rows=200 width=8) (actual time=0.054..3.198 rows=133 loops=1)
Index Cond: (id = ANY ('{...}'::bigint[]))- Cold: 4.1 s (2,685 read, 1,386 dirtied). Warm: 779 ms (589 read, 1,082 dirtied). Almost all of it is index maintenance I/O on
namespaces. The row lookup takes 3 ms. - Plan: https://console.postgres.ai/gitlab/projects/gitlab-production-main/sessions/66985/commands/171796
Known limitations
- A subgroup that scheduled its own deletion is restored like any other descendant when its parent cancels. Keeping its schedule, and stopping propagation there, is a follow-up noted on #633441 (closed).
- Descendants in a non-propagated state (for example
deletion_in_progress) are skipped, along with their subtrees.
References
- #633441 (closed)
- ADR 003: State propagation model
- ADR 006: State preservation
- Blocks #599305 (closed)
How to set up and validate locally
bin/rspec spec/models/concerns/namespaces/stateful/state_precedence_spec.rb
bin/rspec spec/models/concerns/namespaces/stateful_spec.rb \
spec/services/namespaces/state_propagation_service_spec.rb \
spec/workers/namespaces/state_propagation_worker_spec.rbMR acceptance checklist
Evaluate this MR against the MR acceptance checklist. It helps you analyze changes to reduce risks in quality, performance, reliability, security, and maintainability.
Related to #633441 (closed)