Ordered window starts two full-source reads after a signalling refetch with no unexpressible page end #2268
Send to my agentSummary#
After one signalling refetch, a joined ordered window with an unindexed source starts two full-source reads. The ordered-work oracle allows none for this history: no page in it ends on a null or NaN rank, so the documented full-source fallback does not apply.
This is a different counterexample from #2221. That one needs a NaN rank, an anti-joined filter, and a descending window, and it exceeds a bound of one. This one has a single row with rank 0, a joined filter, an ascending window, and exceeds a bound of zero.
The failure is on main. A randomized CI campaign found it while testing an unrelated pull request (#2103, which does not change packages/db).
Law#
packages/db/tests/query/ordered-work-oracle.property.test.ts, in the eager indexed ordered windows lane: after a signalling refetch, if no page end became unexpressible, the loader starts no new full-source request.
if (signalsRefetch && unexpressiblePageEnds === unexpressibleBefore) {
expect(
fullSourceRequests,
`full-source requests after a signalling refetch at step ${step}`,
).toBe(fullSourceBefore)
}
Reproduce#
On main at 7e46f81ad, from packages/db:
TANSTACK_DB_ORACLE_SEED=2146201763 npx vitest run --coverage.enabled=false \
tests/query/ordered-work-oracle.property.test.ts \
-t "eager indexed ordered windows keeps rows and per-change work bounded for a random or replayed seed"
Result, every run:
AssertionError: full-source requests after a signalling refetch at step 1: expected 2 to be +0
Shrunk counterexample:
{"path":"on-demand","provider":"refetch-signal","latency":"immediate","cancellation":"cooperative","peer":false,"filter":"joined","via":"marker","indexed":false,"direction":"asc","nulls":"first","terms":1,"limit":1,"offset":0,"initial":[],"history":[{"type":"upsert","id":0,"rank":0,"eligible":false}]}
The cause is not diagnosed. The seed was first seen in CI job Test (db-1) for #2103 at cb9db1da1, which reproduced the same counterexample.
Sign in to join this conversation.
A second seed reaches the same law, probably through the same cause. CI found it on #2103 (
Test (db-1)at9d0ee7234), where it timed out at 5 s because shrinking took about 9 s.On
mainat7e46f81ad, frompackages/db:Shrunk counterexample: the same joined, ascending,
nulls: first, unindexed window with an ineligible upsert, here with anullrank and a window move:Under the default 5 s test timeout, this seed reports a timeout instead of the assertion.