Comment 3 for bug 1269842

Correct me if I'm wrong, but from what I understand the crash is happening only on the master node. And it is happening at the stage of writeset creation, so it does not look like https://bugs.launchpad.net/codership-mysql/+bug/1254179.

Signal 7 is "bus error" and must be caused by a bug in how Galera 2.x manages mmapped output buffer and is definitely related to sizes. Usually signal 7 is a result of mmapped file not fitting on disk. How much free disk space is in the mysql data directory? Note that with such big transaction you need more than double the disk space, since transactional IO cache will be spilled to disk as well.

However,

1) ATM there is no effective check for transaction exceeding wsrep_max_ws_size.
2) Transaction writeset contains not only row data, but also all unique keys which take additional arbitrary space
3) Galera transaction limit is 2GB. In Galera 2.x it is checked at replication time, but the crash happens well before.
4) 20M rows - no replication will thank you for that. pt-archiver tool is the way to truncate tables. Anyway, with Galera 2.x it will likely require several gigs of spare RAM to process such writeset, which you likely don't have. And if you don't have swap it will get even uglier. So if you insist on going that way, you should be using Galera 3.x.