-
Type:
Epic
-
Resolution: Unresolved
-
Priority:
Major - P3
-
None
-
Affects Version/s: None
-
Component/s: Cursors, Transactions
-
None
-
Storage Engines - Foundations
-
133.054
-
None
-
None
-
Affected readers during disaggregated stepdown
Outcome
Some readers return WT_ROLLBACK during disaggregated stepdown. When the server cannot handle the rollback transparently, customers see a failed read and lose read availability. This breaches the stepdown contract that role transitions should be transparent to customers. The outcome is for affected reads that are intended to survive stepdown to complete without this failure, with their required consistency preserved.
Why the outcome is not met today
Two WT paths can roll back affected readers, including transactions that have only performed reads:
- Case A: An untimestamped snapshot established on the leader cannot be served consistently by the follower checkpoint view after stepdown.
- Case B: A live stable-table open chooses leader mode, races stepdown and is refused (returns WT_ROLLBACK) after the node becomes a follower.
Plan
- Case B, WT-19018: Complete the WT-side retry under the new role. The candidate fix is in a draft PR and has passed WT Evergreen patch testing.
- Case A, WT-19050: Investigate Alex Gorrod's proposal to retain this node's live stable view, with local transaction IDs intact, for eligible pre-stepdown snapshots.
- Validate both cases with WT tests and the server, including reads that have already returned results. A restart must not silently change their snapshot.
Completion criteria
Affected reads covered by the stepdown contract survive both cases without a customer-visible failure or a change in required consistency. Completing the Case A investigation alone does not complete the epic.
For more detail, see the design document.