-
Type:
Task
-
Resolution: Unresolved
-
Priority:
Major - P3
-
None
-
Affects Version/s: None
-
Component/s: None
-
None
-
DB Integration & Observability
-
None
-
None
-
None
-
None
-
None
-
None
-
None
Summary
The sharded whole-cluster EOT change stream passthrough suites (introduced in SERVER-137666) fail with repeated Location8027900 errors on disaggregated storage, across several distinct tests. The suites were removed from SERVER-137666 pending this investigation.
Symptom
"attempting to continue transaction that was not started lsid: - - txnNumber: N", code 8027900 (Location8027900).
A shard receives an autocommit: false continuation for an implicit transaction it never saw started. Observed on at least three distinct tests with different txnNumbers:* jstests/change_streams/change_stream_resume_point_absent_fails_fast.js (txnNumber 0)
- jstests/change_streams/filter_excludes_resume_point.js:19 (txnNumber 0)
- jstests/change_streams/resume_from_high_water_mark_token.js (txnNumber 2)
Context
- Suite shape: ShardedClusterFixture (2 shards, 3 mongos), DisaggShardedClusterFixture on disagg, wrapCRUDinTransactions txn override, whole-cluster upconversion (implicit_whole_cluster_changestreams.js — watch aggregates execute against admin via allChangesForCluster: true), plus showEndOfTransaction: "required" stamping.
- The failures were observed ONLY on the disagg variants. On plain-sharded rbe runs the equivalent tasks failed in the 'fetch remote test results' infrastructure step, so whether plain-sharded whole-cluster exhibits the same test failures is UNKNOWN (masked).
- The whole-db shape (same txn machinery, no whole-cluster upconversion) is green in the same patch builds — the whole-cluster upconversion's admin-db watch path is the differentiating factor.
Open questions
- Exact failing statement path: the test's write continuation via mongos, or the upconverted watch aggregate on admin joining an implicit transaction started on a different database's primary shard?
- Is the failure disagg-routing-specific (SLS cell routing) or a general cross-shard implicit-txn continuation gap that plain-sharded runs would also hit once the infra masking is resolved?
- Does the txn override (network_error_and_txn_override.js) need to avoid implicit transactions spanning databases, or does the mongos/shard transaction continuation need fixing?
- is related to
-
SERVER-137666 Create sharded passthrough test for broader endOfTransaction coverage
-
- In Progress
-