Ordered refetch retains overlapping first-page requests after the backend becomes empty

Offen
#2,414 0 Kommentare 0 Reaktionen 0 zugewiesene Personen Auf GitHub ansehen

Maintainer antworten meist innerhalb von 1 Tag

Dieses Issue hat noch niemand übernommen.

Bewertung

Schwierigkeit
4/5
Geschätzter Aufwand
3-5 Tage
Anfängerfreundlichkeit
48/100
Issue-Typ
Bug
Klarheit
Klar beschrieben
Aktivitätsstatus
Aktiv
Tech-Stack
typescript
Bereich
database, testing

Rechercherichtung

Start with the replay command in the issue and the ordered request contract in packages/db/src/query/live/ARCHITECTURE.md, then inspect finishOrderedPrefixRepair and retireSettledFiniteAcquisitions in packages/db/src/query/live/ordered-source-loader.ts. Preserve the reduced history through the shared model, driver, and comparison, and extend acquisition-retirement histories for empty backends and overlapping repair prefixes. Done means the replay passes after the repair, fails on unchanged production, and rejects the original defect as a mutant while preserving the listed cleanup behaviors.

Vom Indexierungsmodell aus dem Issue-Text verfasst.

Beschreibung

Summary

The random query-db.refetch-ordered-prefix campaign found an unchanged-backend refetch that sends two overlapping first-page requests instead of at most one. Deleting every backend row is enough to reach it in the reduced history.

Observed on main at 3f05109f60eee7222118b3d838508c3ed046ad7d in Release / Test (collections), attempt 1. The same seed and shrink path reproduce locally on unchanged production. The random campaign may pass with another seed; that does not establish that this history is fixed.

This failure prevented that run's dependent Release job from starting. The maintainer restarted the workflow and requested that this counterexample be tracked separately from release recovery. No production code, oracle expectations, comparisons, or run budgets were changed.

Reduced history

Property: query-db.refetch-ordered-prefix

seed: 1067726211
path: 41:0:1:2:2:2:3:2:2:7:8:8:8
{
  "scenario": {
    "encoding": "cursor",
    "functionKey": false,
    "terms": 1,
    "orderKey": "rank",
    "limit": 1,
    "offset": 1
  },
  "initial": [
    { "id": 2, "rank": 0, "marked": false },
    { "id": 4, "rank": 0, "marked": false }
  ],
  "history": [
    { "type": "delete", "id": 4 },
    { "type": "delete", "id": 2 },
    { "type": "refetch" }
  ]
}

pageCap is undefined: the provider is uncapped. With one order term the fixture maps rank to rank * 100 + id, so the source ranks are 2 and 4, not tied. Both rows are unmarked, so the LEFT-joined marker filter leaves the public window empty while the source loader pages.

The shared driver appends another refetch and repeats unchanged-backend refetches and invalidation before its final comparisons. The failed checkpoint is the repeated-refetch request-work comparison, not the window or source-coverage comparison:

a refetch of an unchanged backend started 2 first-page requests; expected at most 1

Observed request work

Temporary logging at the existing failing checkpoint recorded these two requests during a repeated refetch after the backend became empty:

orderBy: rank desc, limit: 2, offset: 0, no cursor
orderBy: rank desc, limit: 1, offset: 0, no cursor

Acquisitions observed earlier in the same run:

  1. Ordered prefix, limit 2.
  2. Boundary tie, rank == 2.
  3. Cursor suffix, limit 1, offset 2, rank < 2.
  4. Repair prefix, limit 2, refetch: true.
  5. Replacement repair prefix, limit 2, refetch: true.
  6. Repair first page, limit 1, offset 0, refetch: true.

This suggests redundant prefix acquisitions survive bounded-repair completion. finishOrderedPrefixRepair / retireSettledFiniteAcquisitions in packages/db/src/query/live/ordered-source-loader.ts are candidate repair locations: completion keeps acquisitions belonging to the current repair chain. A causal fix has not been implemented or verified. Temporary diagnostic logging was removed after capturing the requests.

Replay

From the repository root at the recorded commit, with dependencies installed and Query Collection dependencies built:

TANSTACK_DB_ORACLE_PROPERTY=query-db.refetch-ordered-prefix \
TANSTACK_DB_ORACLE_SEED=1067726211 \
TANSTACK_DB_ORACLE_PATH=41:0:1:2:2:2:3:2:2:7:8:8:8 \
pnpm --config.verifyDepsBeforeRun=false \
  --filter @tanstack/query-db-collection exec vitest run \
  tests/refetch-ordered-prefix-oracle.test.ts \
  --maxWorkers=1 --fsModuleCache=false \
  --typecheck.enabled=false --coverage.enabled=false

Local result: the replay fails at the same request-work comparison in approximately 11 seconds; 33 other tests pass, 2 are expected failures, and 9 are skipped by replay selection.

Contract and follow-up

The existing oracle's neighboring work law requires an unchanged-backend refetch or invalidation to start at most one first-page request for this window's predicate and order. Its authority is the ordered request/acquisition contract in packages/db/src/query/live/ARCHITECTURE.md, “Ordered requests, continuation, and recovery.” The recorded history does not fail public window or source-coverage judgments; broader effects have not been ruled out.

Preserve this history through the shared model, driver, and comparison. Extend generated acquisition-retirement histories to exercise empty backends and overlapping repair prefixes. Any cleanup repair must preserve covering demand, whole-chain settlement, truncate replay, and reentrant release behavior. Demonstrate RED on unchanged production, GREEN after repair, and rejection of the original defect as a mutant.

Related, but different observations:

  • #2186: broader ordered-refetch oracle follow-up.
  • #2191: acquisition-ledger modeling; its stated limits exclude same-epoch prefix supersession and tie/refill sub-steps.
  • #2268: two full-source reads after a signalling refetch, rather than two bounded first-page requests.
  • #2335: source-coverage failure with two order terms, rather than this request-work failure.
Vorherrschende Sprache
TypeScript
Sterne
4k
Forks
286
Ø Merge
1 T. 1 Std.
Gemergte PRs (30 T.)
104

Entwicklungsumgebung

Erste Schritte

  1. Lesen Sie das ganze Issue und danach den Beitragsleitfaden des Projekts.
  2. Schreiben Sie ins Issue, dass Sie es übernehmen — das erspart doppelte Arbeit.
  3. Forken Sie das Repository und arbeiten Sie in einem Branch.
  4. Öffnen Sie einen Pull Request, der die Issue-Nummer nennt.

Mehr aus TanStack/db

Alle Issues in TanStack/db

Ähnliche Issues

Weitere Issues zu TypeScript

Neue Issues direkt in Ihr Postfach

Eine kurze Übersicht über anfängerfreundliche GitHub-Issues.