-
Type:
Task
-
Resolution: Unresolved
-
Priority:
Major - P3
-
None
-
Affects Version/s: None
-
Component/s: Timestamps, Transactions
-
None
-
Storage Engines - Transactions
-
219.736
-
SE Transactions - 2026-12-18
-
None
Problem
With preserve_prepared enabled, __txn_validate_durable_timestamp rejects a prepared transaction whose durable timestamp is less than or equal to its prepare timestamp (added by WT-17377). In disaggregated storage this restriction is believed to be unnecessary.
Rationale
- A follower already rejects any prepare, commit or durable timestamp at or below the last picked-up checkpoint timestamp (__txn_check_follower_checkpoint_ts,
WT-18878for commit and durable,WT-18943for prepare). - On the leader, the durable timestamp must be after the stable timestamp, and the leader's checkpoint timestamp does not exceed stable.
- The stricter "durable timestamp strictly after the last checkpoint timestamp" assert for prepared-id transactions (
WT-17459) is therefore subsumed. - These checks keep the timestamps away from the checkpoint boundary. They do not by themselves require durable to differ from prepare. The justification for removing the durable_ts <= prepare_timestamp check is that
WT-17377only states a general rationale (keeping the prepare and durable moments distinct), and no code was found that depends on the gap: with preserve_prepared a prepared cell is identified by its prepared id, not by a timestamp difference.
Scope
The check is not gated on disaggregated storage. It applies to any connection with preserve_prepared, and test_prepare50 covers it on a plain table: URI. Decide whether to remove it only for disaggregated storage or entirely.
Proposed change
- Remove (or gate off for disaggregated storage) the durable_ts <= prepare_timestamp check in src/txn/txn_timestamp.c.
- Remove the redundant assert on the last checkpoint timestamp in __txn_set_durable_timestamp and update its comment.
- Update test_prepare50 to match the new behaviour.
- Change test/checkpoint to use the same timestamp for the prepare, commit and durable timestamps of a prepared transaction. The worker threads currently commit with durable at the prepare timestamp plus 2 (workers.c), and the prepare-discover thread uses durable at commit plus 1 (checkpointer.c).
- Change test/format to use the same timestamp for prepare, commit and durable. The predictable-replay prepare timestamp selection (replay.c) currently picks a prepare timestamp strictly below the commit timestamp, and skips preparing when none is available, to satisfy the durable > prepare rule. Update that logic and its comment.
- is related to
-
WT-18878 Reject follower commits at or below the adopted checkpoint timestamp
-
- Closed
-
-
WT-18943 Reject follower prepares at or below the adopted checkpoint timestamp
-
- Closed
-
-
WT-17377 Enforce that a prepared transaction's durable timestamp is greater than its prepare timestamp
-
- Closed
-
-
WT-17459 test/format (disagg.mode=switch) WiredTiger assertion failed: 'upd->prepare_state != (uint8_t)1 && upd->prepare_state != (uint8_t)2'
-
- Closed
-
- related to
-
WT-19105 Claiming a recovered prepared transaction with debug table logging leaves its transaction ID in the global table
-
- Closed
-