Skip to content- Sources: primary, discussion
- Summary: Since 10.5, freeing undo and B-tree pages ticked freed_page_clock, which switched on LRU promotion under one global buf_pool.mutex even when data fit in memory. The fix sets a PROMOTE bit atomically at page access and lets sweeps that already hold the mutex move flagged blocks. On one test 11.8 rose from 20,326 to 57,052 QPS, about 7 to 9 percent faster than 10.4.34.
- Why it matters: MariaDB users on 10.5 or later with in-memory reads gain from the fixed 11.8 and 12.3 releases, and innodb_old_blocks_time now has one-second resolution.