Ordered loader: a page-capping provider loses a moved row from a joined window after refetch #2183
Send to my agentSummary#
With a provider that returns fewer rows than a request's limit (a page cap), an on-demand joined ordered window can lose a row after a refetch. The row's order changed, and the loader's own repair drops it. A short reply does not prove that the source is exhausted (glossary), so the loader must keep reading.
Reproduction#
The pinned witness is a capped provider keeps a moved row in the window after a refetch (known loader defect) in packages/query-db-collection/tests/refetch-ordered-prefix-oracle.test.ts (branch fix-query-db-refetch-ordered-gap, PR #2154). It is an it.fails test today.
- Provider: cursor encoding,
pageCap: 2. - Query:
rank desc,offset 1,limit 2, LEFT-joined marker filter. - Initial rows: id 0 (rank 1, marked), id 1 (rank 0, unmarked).
- History:
- Mark id 7.
- Upsert id 2 at rank 1.
- Upsert id 7 at rank 0.
- Refetch.
- Delete id 3.
- Invalidate.
- Upsert id 3 at rank 1.
- Upsert id 0 at rank 0.
Failure: window after refetch at step 3 was []; expected [7].
Scope#
This reproduces on main, through the loader's own repair. It is separate from #2136. The query-db refetch oracle's generated campaigns use an uncapped provider because of this defect. A page-capped trace together with a peer live query also fails (found while #2154 was reviewed). That trace fails the same way with the loader's first-page skip disabled.
Suggested witness and owner#
- Owner: the ordered loader. The repair and continuation path must treat a short page as "not exhausted".
- Witness: flip the pinned
it.failstest, then addpageCapto the query-db refetch oracle's generated scenario.
This issue was written by Isaac.
Sign in to join this conversation.