-
Type:
Improvement
-
Resolution: Unresolved
-
Priority:
Major - P3
-
None
-
Affects Version/s: None
-
Component/s: Cursors
-
None
-
Storage Engines - Foundations
-
202.973
-
SE Foundations - 2026-11-24
-
5
Problem. During an asynchronous step-down, a layered cursor operation can start while the node is still the leader and then need to open its stable table after the step-down has completed. That open is refused, and the layered cursor returns WT_ROLLBACK to the application (see WT-18179).
Nothing the operation did causes the rollback: the node's role changed underneath it. It also hits operations that are otherwise never rolled back this way, such as autocommit reads and reads with a read timestamp. Each affected operation fails back to the application, which has to retry it, so reads on the node are unavailable until those retries succeed.
The window is wider than it looks. After a step-up, every stable table handle is opened fresh, and idle handles are closed by sweep, so a step-down that follows soon after a step-up can send many cursors into this race at once and cause a burst of rollbacks.
Why WiredTiger should handle it. The condition is transient and already resolved by the time the open is refused: the node's new role is known at that point. Retrying the operation inside WiredTiger is straightforward and keeps the race invisible to the application.
Outcome. A layered cursor operation that races a step-down completes under the node's new role instead of returning WT_ROLLBACK. Transactions whose snapshot the follower cannot serve consistently are still rolled back, as today. No new error codes reach the application from layered cursor operations. A test reproduces the race deterministically and checks that the operation succeeds.
- is depended on by
-
SERVER-139349 Enable stepdowns in disagg_concurrency_sharded_terminate_kill_primary_with_balancer
-
- In Progress
-