Performance Change: MariaDB 12 shows reduced throughput vs 11 on cached write workloads

Initial performance change detection testing shows that MariaDB 12 exhibits lower throughput than MariaDB 11 during long-running cached write workloads. The behavior appears to be related to redo log processing.

This was observed using TAF with identical configurations and environments across both versions.

To confirm that the regression was tied specifically to the synchronous commit path, OLTP_RW tests were repeated with innodb_flush_log_at_trx_commit=2. Under commit=2, MariaDB 12 not only removed the regression but exceeded MariaDB 11 performance. This demonstrates that the slowdown in MariaDB 12 is not due to SQL execution, optimizer behavior, or general engine overhead. The regression is specifically tied to the commit=1 durability path, where MariaDB 12 performs more redo and checkpoint work per transaction than MariaDB 11.

-------------------------------------------------------------------------------------------------------------+

--------------------------------------------------------- -----------------------------------------------------+

Perf Data: https://drive.google.com/file/d/1Neo30XuIYBhl5tE4LfxQvvHzPLzh8xzj/view?usp=drive_link

Perf data was generated by running 128 threads for 50 minute, and taking a 35 second slice around the 45 min mark of the iteration.

I’m happy to work with someone in PLC to help confirm and isolate the issue.