Currently, the round_to_oldest functionality observes the global oldest timestamp after the transaction's snapshot has already been generated. This is problematic because the oldest timestamp that is observed represents a different, later state of uncommitted transactions than what was captured by __wt_txn_get_snapshot(). This can cause an undesirable slicing between what is visible due to snapshot isolation and what is visible due to the read timestamp, if some transactions happen to commit in between getting the snapshot and rounding the read timestamp to the oldest timestamp.
- is depended on by
-
SERVER-34590 oplog visibility issues with round_to_oldest
- Closed