Jepsen: MariaDB Galera Cluster 12.1.2
3 hours ago
- #Database Consistency
- #Galera Cluster
- #MariaDB
- MariaDB Galera Cluster claims to offer strong consistency but fails to provide Snapshot Isolation or Repeatable Read.
- Under recommended settings, Galera Cluster loses committed transactions during coordinated node failures.
- Even with safer settings, Galera Cluster occasionally loses committed writes during process crashes and network partitions.
- Galera Cluster exhibits Lost Update (P4) and Stale Read anomalies even in healthy clusters without faults.
- MariaDB's documentation is misleading, claiming 'no lost transactions' and 'instant replication', which contradicts observed behavior.
- Users are advised to set `innodb_flush_log_at_trx_commit=1` to reduce write loss but should still expect anomalies.
- Four major issues were identified and reported to MariaDB: MDEV-38974, MDEV-38976, MDEV-38977, and MDEV-38999.
- Future work could explore additional failure scenarios, such as clock skew, disk faults, and predicate-based anomalies.