Ordered-replay oracle: terminal error after a failed settle following truncate markers #2224
Send to my agentSummary#
The random campaign of packages/db/tests/query/load-subset-replay-refinement-oracle.test.ts › "Ordered windows across truncate replays › keeps the ordered window under load and release faults" found a counterexample on main at 30175be48 (#2194). The failure is deterministic when the seed is replayed.
Replay#
cd packages/db
TANSTACK_DB_ORACLE_SEED=1817832416 \
TANSTACK_DB_ORACLE_PATH=115:1:2:2:2 \
TANSTACK_DB_ORACLE_PROPERTY=replay.ordered-window-faults \
pnpm vitest run tests/query/load-subset-replay-refinement-oracle.test.ts -t "release faults"
Counterexample (shrunk)#
[
[{ "type": "hold" }, { "type": "truncateMarkers" }, { "type": "settle", "ok": false }],
{ "direction": "asc", "offset": 0, "limit": 1, "ties": false, "peer": false,
"joined": "none", "markerMode": "on-demand", "includes": true }
]
AssertionError: convergence: terminal error: expected false to be true
at checkOrderedReplay (tests/query/load-subset-replay-refinement-oracle.test.ts:1875)
The shown window is in an error state, but the oracle's model has no failed load that explains it (fatalLoadFailed() is false). Either production shows a terminal error that the history does not justify, or the model is missing a case where a failed settle after truncate markers is fatal.
Where it was seen#
CI for #2219 (job Test (db-1), run 39726325795). #2219 does not change @tanstack/db; CI tests that PR merged with main. The replay above fails with main's packages/db.
This issue was written by Isaac.
KyleAMathews(GitHub profile, opens in a new tab) closed this
🎉 The fix for this issue has been released!
Fixing PRs: https://github.com/TanStack/db/pull/2231, https://github.com/TanStack/db/pull/2266
Backfilled notification for published release candidates.
Sign in to join this conversation.
Fixed by #2229 (merged as
1640de834).On
mainat1640de834, the shrunk trace from this issue passes: hold,truncateMarkers, then a failedsettle, on an ascendinglimit: 1window withjoined: "none",markerMode: "on-demand", andincludes: true. A second random-campaign seed that found the same class also passes now:TANSTACK_DB_ORACLE_SEED=1476128014 TANSTACK_DB_ORACLE_PATH=191:3:9:4:4:4:4 TANSTACK_DB_ORACLE_PROPERTY=replay.ordered-window-faults.This turned out to be a gap in the model, not a production defect. The failed load is new include demand issued during the marker replay, so the terminal error is fatal. #2229's model now classifies it that way, through its
new-demand-failed-during-marker-replaystate.A follow-up PR keeps this exact trace as a named replay in
load-subset-replay-refinement-oracle.test.ts, withterminal: true.#2239 reports the same trace and is closed as a duplicate.