Details
-
Bug
-
Status: Open (View Workflow)
-
Major
-
Resolution: Unresolved
-
None
-
None
-
None
Description
Version: 11.4 branch at 095cfb5e468e (2026-11-11), DuckDB v1.5.5, built with storage/duckdb/build.sh. Related: MDEV-40957 (read-free DML on the slave).
Setup: primary MariaDB 10.11.14 (binlog_format=ROW, binlog_row_image=FULL) running sysbench oltp_read_write at 400 tps on 8 tables. On the replica the same tables exist as ENGINE=DuckDB but are EMPTY (a failed initial load on our side).
Steps:
1. CHANGE MASTER TO ... (position of the dump); START SLAVE;
2. Let it run a few minutes.
Result: replication keeps running with Last_SQL_Errno 0. Duckdb_rows_update_in_batch and Duckdb_rows_delete_in_batch grow (106,934 updates and 53,467 deletes in our case), although none of those rows exist on the replica. The tables end up holding only the rows inserted after the start, and nothing reports that the replica has diverged.
Expected: an Update_rows or Delete_rows event for a row that is not there stops replication with HA_ERR_KEY_NOT_FOUND (1032), as with InnoDB, or at least is counted or logged, so a replica that diverged does not keep answering queries as if it were in sync.
Note: i understand read-free apply skips the lookup on purpose (rnd_pos_by_record returns 0 for the applier). The cost is that slave_exec_mode=STRICT no longer catches divergence for these tables.