Details
-
Bug
-
Status: Closed (View Workflow)
-
Blocker
-
Resolution: Fixed
-
10.3.36, 10.4.26, 10.6.11
-
None
-
Centos 7.9
Description
Hello,
We encounter a problem while upgrading from 10.4.18 to 10.4.26.
All was OK on "single" servers for dev, integration but when we upgrade pre-production servers (cluster) we get a big problem.
Our clusters are built with one "master" machine and a "slave" machine.
The master works well and all work is done, but on the replica the slave process is blocked with locks.
We use some tables with sequences to have an "unique" numbering scheme for 2 or 3 tables and the replica block when receiving the data for the sequence and a corresponding table.
One thing made me wonder, on the master the sequence locks are "MDL SHARED_WRITE" and on the replica it locks with "MDL_EXCLUSIVE", the depending table is locked with "MDL SHARED WRITE".
I can kill the locked process but the replica locks again for the next insert.
We can not upgrade the production servers as many tables are using sequences.
We do not have any problems with that kind of cluster since the version 10.2.
The binlog is using "mixed" format.
Any help will be welcome ...
PS : the production servers are working at 10000 to 45000 transactions/ minute and the replica's lag is between 0 and 2 seconds.
JP P
Attachments
Issue Links
- causes
-
MDEV-31779 Server crash in Rows_log_event::update_sequence upon replaying binary log
- Closed
- relates to
-
MDEV-30077 Replication locked after migration of 1 server from 10.4.18 to 10.4.25
- Closed