Details
-
Bug
-
Status: Open (View Workflow)
-
Critical
-
Resolution: Unresolved
-
N/A
-
None
-
Related to performance
Description
In our tests, the implementation of BACKUP SERVER TO (MDEV-14992, copying to a mounted file system) can be slower than mariadb-backup --backup when block cloning is not available. (Backing up within XFS or btrfs or similar will obviously give an advantage over the legacy implementation, which is missing MDEV-23947.)
A possible explanation is that the new implementation will defer the copying of the InnoDB write-ahead log towards the end of the backup, while the dedicated log_copying_thread() in the older implementation starts copying the log right away. I have implemented a prototype (fixed up) of more eager log copying in the multi-threaded BACKUP SERVER TO operation. The test backup.backup_innodb is failing on multiple targets, and the implementation currently only covers the memory-mapped log writing path (MDEV-25124). A few more days of work is needed to implement this before any performance comparison can be conducted to validate this approach.
Note: The performance of streaming BACKUP SERVER WITH (MDEV-38362) appears to be competitive with mariadb-backup even when the log streaming is initiated towards the end.
This performance improvement is filed as a separate bug so that we could release what we currently have without a delay.
Attachments
Issue Links
- is caused by
-
MDEV-14992 BACKUP SERVER to mounted file system
-
- In Progress
-