On Wed, Jan 18, 2023 at 4:30 PM Marko Mäkelä <marko.makela@mariadb.com> wrote:
It should be worth noting that starting with MariaDB 10.5, there will no longer be writes of garbage pages that were freed from the underlying file, for example, as part of DROP INDEX, TRUNCATE TABLE, or a massive DELETE.
In https://jira.mariadb.org/browse/MDEV-19895 I would welcome constructive feedback on how to set sane default values of some parameters, based on a small number of parameters, say, the available size of memory, CPU cores, storage, and the size of the database.
May or may not be directly related to this, but with 10.5+ I see the checkpoint age exhibiting a sawtooth pattern - nothing gets flushed until it gets near the log capacity, then there is a huge burst of writes and it crashes down to a sensible level. Prior to 10.5, the checkpoint age remains low unless the workload is overruning innodb_io_capacity. I get the motivation to reduce flushing, but if the checkpoint age is already close to the high water mark and a sustained high volume of transactions hits, then the server has nowhere to go because it has already used up all of it's scope for absorbing the shock into the redo log and everything grinds to a halt. Is there a way to restore the old behaviour? The new behaviour seems to be in all 10.5+ versions I looked at.