InnoDB enables an ‘adaptive’ flushing mode by default, which attempts to ensure a steady flow of page flushing and avoid write bursts on log checkpoint.

There is a non-default option innodb_adaptive_flushing=OFF that is still based on waking up the page cleaner thread at least once per second. I think that we should offer a purely event-driven flushing option that would ignore innodb_io_capacity altogether and perform best effort for the buffer pool to remain between innodb_max_dirty_pages_pct_lwm and innodb_max_dirty_pages_pct.

I implemented a crude prototype of this.