On Sunday 10 October 2021 23:58:12 Otto Kekäläinen wrote: > > The problem is in the ibdata1 file (about 450MB). Deleted other database > > directories and it still crashes, deleted ibdata1 and it runs. > > > > How to bisect mariadb from git? Tried: > > $ git bisect good mariadb-10.3.29 > > $ git bisect bad mariadb-10.3.31 > > the build process showed version 10.2 so I aborted it. > > > > Checked out mariadb-10.3.30 but dpkg-buildpackage failed with: > > dh_install: mariadb-plugin-cassandra missing files: > > etc/mysql/conf.d/cassandra.cnf > > Some dependency was missing and Cassandra was not built. Note that the > upstream repository is not identical to the one in Debian regarding > the contents of debian/ directory. MariaDB builds without a cache take > 30 mins each and there are all kinds of things going on. Doing bisect > (fully correctly) on MariaDB is hard even for experienced developers. > Your time is probably better spent doing some other kind of debugging. > > I recommend that you file a bug about this upstream, and try to attach > relevant info from the error log, maybe a strace output etc. Upstream > devs will guide you on what to debug next. > > One thing you could also try is to start the server with 10.3.29 and > ensure that you have a clean shutdown (SET GLOBAL > innodb_fast_shutdown=0; SHUTDOWN) and only after that start with the > new 10.3.31 binary.
With fast shutdown disabled, mysqld (10.3.29) seems to be stuck in an infinite loop. 100% CPU usage, no I/O for an hour until I killed it. Even dropping all the databases did not help - killed it after it has been running for 5 minutes. Deleted ibdata1 (and ib_logfile0, ib_logfile1), then shutdown ended immediately. Seems that the file structure is corrupted somehow - probably because of a previous bug. One table is also affected by the "ERROR 1118 (42000): Row size too large (> 8126)" bug. When complete SQL dump is restored with a new ibdata1, everything works (upgrade to 10.3.31 and also clean shutdown). -- Ondrej Zary