fix(data): serialize new-link publication races

This commit is contained in:
2026-06-02 07:31:35 +03:00
parent 1d62f69ff0
commit ea714480d3
5 changed files with 105 additions and 10 deletions
@@ -78,7 +78,8 @@
- `da0a306` `fix(data): enforce completed portfolio sessions`
- `a28b75d` `fix(data): align portfolio mutation lock order`
- `d762ecc` `fix(data): serialize portfolio future reschedules`
- Current fix cycle: `fix(data): lock racing portfolio publications`
- `1d62f69` `fix(data): lock racing portfolio publications`
- Current fix cycle: `fix(data): serialize new-link publication races`
**Files:**
- Create: `tests/GmRelay.Bot.Tests/Web/PortfolioMigrationTests.cs`
@@ -267,6 +268,9 @@ public async Task PublishingDraftCardWithAnyFutureLinkedSession_ShouldFailCommit
[Fact]
public async Task ConcurrentPublishAndFutureReschedule_ShouldNotDeadlockOrCommitInvalidPublicCard()
[Fact]
public async Task ConcurrentNewLinkPublishAndFutureReschedule_ShouldNotCommitInvalidPublicCard()
[Theory]
[InlineData("portfolio_game_sessions")]
[InlineData("portfolio_game_masters")]
@@ -281,7 +285,7 @@ public async Task RequiredParentCascadeDelete_ShouldFailCommitForPublishedCard(s
public async Task ParentCardAndGroupCascadeDeletes_ShouldCommit()
```
The direct-delete, moved-link, invalid publication, and direct parent-cascade scenarios must expect PostgreSQL `23514` at commit. Every selected linked session must be completed with `scheduled_at < now()`: one future link among multiple selected sessions rejects publication. A final future reschedule must atomically unpublish linked public cards while preserving their first `published_at`; `past -> future -> past` in one transaction must leave the card public. Opposing-order batch reschedules must use an advisory test gate plus `pg_blocking_pids` observation with bounded timeouts, complete without card deadlock, and leave both cards private; do not rely on `pg_sleep` timing. The `READ COMMITTED` concurrency scenarios must launch bounded tasks together, cover both publish/delete lock orders, and prove there is no deadlock, write-skew, or invalid public commit. A session-delete versus future-reschedule race must use the common `sessions` then `portfolio_games` lock order, cover both first-session-lock orders through real blocking transactions, and finish with the card private and session deleted. The publish/reschedule race must finish with the future session committed and the card private. The `REPEATABLE READ` scenarios must reject triggered portfolio writes with `0A000`, including both draft-link deletion versus publication commit orders, because a stale snapshot after lock acquisition cannot safely validate the invariant. The parent-card and owning-group cascade scenarios must commit successfully.
The direct-delete, moved-link, invalid publication, and direct parent-cascade scenarios must expect PostgreSQL `23514` at commit. Every selected linked session must be completed with `scheduled_at < now()`: one future link among multiple selected sessions rejects publication. A final future reschedule must atomically unpublish linked public cards while preserving their first `published_at`; `past -> future -> past` in one transaction must leave the card public. Opposing-order batch reschedules must use an advisory test gate plus `pg_blocking_pids` observation with bounded timeouts, complete without card deadlock, and leave both cards private; do not rely on `pg_sleep` timing. The `READ COMMITTED` concurrency scenarios must launch bounded tasks together, cover both publish/delete lock orders, and prove there is no deadlock, write-skew, or invalid public commit. A session-delete versus future-reschedule race must use the common `sessions` then `portfolio_games` lock order, cover both first-session-lock orders through real blocking transactions, and finish with the card private and session deleted. The publish/reschedule races must finish with the future session committed and the card private, including a new-link draft publication forced behind the post-row-lock advisory gate. The `REPEATABLE READ` scenarios must reject triggered portfolio writes with `0A000`, including both draft-link deletion versus publication commit orders, because a stale snapshot after lock acquisition cannot safely validate the invariant. The parent-card and owning-group cascade scenarios must commit successfully.
- [ ] **Step 3: Run the Task 1 tests to verify RED**
@@ -435,6 +439,8 @@ BEGIN
ORDER BY pg.id
FOR UPDATE OF pg;
PERFORM pg_advisory_xact_lock(20260530, 108);
UPDATE portfolio_games pg
SET is_public = false,
updated_at = now()
@@ -508,7 +514,7 @@ CREATE INDEX ix_portfolio_game_reviews_pending
WHERE moderation_status = 'Pending';
```
The deferred constraint triggers retain the link-table `ON DELETE CASCADE` behavior. At transaction commit validators acquire the same transaction-level advisory lock and reject a surviving published card when either required link set is empty or any linked session has `scheduled_at >= now()`. The intentionally global lock is appropriate for low-volume portfolio publication writes: under the application default `READ COMMITTED` isolation level it serializes validation, prevents write-skew across distinct child links, and gives multi-card transactions one lock order. PostgreSQL retains stale snapshots under `REPEATABLE READ` and `SERIALIZABLE`, so the guard rejects every triggered portfolio write at those isolation levels with `0A000`. The deferred future-reschedule trigger re-reads the final session row, skips intermediate future values that end in the past, and for a final future value locks all cards linked to any final-future session in `portfolio_games.id` order before one guarded public-card unpublish update. The lock phase deliberately includes committed drafts so a concurrent draft-to-public publication cannot pass validation against the pre-reschedule session snapshot and commit afterward. This separate global row-lock pass avoids opposing batch order without adding the validator advisory lock before card locks. At `READ COMMITTED`, draft edits, explicit unpublishing, future reschedules, and card or club cascade deletion remain valid. Normal session-deletion handlers use the same `sessions` then `portfolio_games` lock order: explicitly lock the target session row, unpublish linked cards, then delete the session.
The deferred constraint triggers retain the link-table `ON DELETE CASCADE` behavior. At transaction commit validators acquire the same transaction-level advisory lock and reject a surviving published card when either required link set is empty or any linked session has `scheduled_at >= now()`. The intentionally global lock is appropriate for low-volume portfolio publication writes: under the application default `READ COMMITTED` isolation level it serializes validation, prevents write-skew across distinct child links, and gives multi-card transactions one lock order. PostgreSQL retains stale snapshots under `REPEATABLE READ` and `SERIALIZABLE`, so the guard rejects every triggered portfolio write at those isolation levels with `0A000`. The deferred future-reschedule trigger re-reads the final session row, skips intermediate future values that end in the past, and for a final future value locks all cards linked to any final-future session in `portfolio_games.id` order. It then acquires the publication advisory lock and runs one guarded public-card unpublish update with a fresh `READ COMMITTED` statement snapshot. The row-lock phase deliberately includes committed drafts so a concurrent draft-to-public publication cannot pass validation against the pre-reschedule session snapshot and commit afterward; the post-row-lock advisory phase also serializes a previously invisible concurrent link-add publication without moving the advisory lock above card locks. At `READ COMMITTED`, draft edits, explicit unpublishing, future reschedules, and card or club cascade deletion remain valid. Normal session-deletion handlers use the same `sessions` then `portfolio_games` lock order: explicitly lock the target session row, unpublish linked cards, then delete the session.
- [ ] **Step 5: Lock sessions before explicitly unpublishing linked cards in both session-deletion handlers**
@@ -554,7 +560,7 @@ Run:
dotnet test tests/GmRelay.Bot.Tests/GmRelay.Bot.Tests.csproj --filter "FullyQualifiedName~PortfolioMigrationTests|FullyQualifiedName~PortfolioMigrationPostgresTests|FullyQualifiedName~PortfolioSessionDeletionSourceTests|FullyQualifiedName~PortfolioSchemaGateSourceTests"
```
Expected: PASS, including PostgreSQL 17 migration application, rejected direct required-link deletes, rejected moved links and session/player cascades, rejected publication with any future linked session, automatic unpublish with preserved `published_at` after final future reschedule, preserved public state after `past -> future -> past`, opposing-order batch reschedules without card deadlock, bounded `READ COMMITTED` publish/delete in both commit orders, publish/reschedule races, session-delete/reschedule serialization in both first-lock orders, and distinct-link deletion without deadlock, write-skew, or invalid public commit, rejected `REPEATABLE READ` triggered writes including both draft-delete versus publish commit orders, successful parent-card and owning-group cascades, Discord identity scoping, and Compose/Aspire HTTP health gating with a non-proxied bot endpoint and matching `gmrelaydb` resource name.
Expected: PASS, including PostgreSQL 17 migration application, rejected direct required-link deletes, rejected moved links and session/player cascades, rejected publication with any future linked session, automatic unpublish with preserved `published_at` after final future reschedule, preserved public state after `past -> future -> past`, opposing-order batch reschedules without card deadlock, bounded `READ COMMITTED` publish/delete in both commit orders, existing-link and new-link publish/reschedule races, session-delete/reschedule serialization in both first-lock orders, and distinct-link deletion without deadlock, write-skew, or invalid public commit, rejected `REPEATABLE READ` triggered writes including both draft-delete versus publish commit orders, successful parent-card and owning-group cascades, Discord identity scoping, and Compose/Aspire HTTP health gating with a non-proxied bot endpoint and matching `gmrelaydb` resource name.
- [ ] **Step 7: Commit**