Jepsen's analysis of MariaDB Galera Cluster 12.1.2 uncovered multiple serious consistency and durability bugs. With MariaDB's recommended configuration (innodb_flush_log_at_trx_commit=0), coordinated node crashes caused committed transactions to be silently lost. Even with the safer setting of 1, write loss still occurred under process crashes and network partitions. Beyond data loss, healthy clusters exhibited Lost Update (P4) and Stale Reads under normal operation—violating both Snapshot Isolation and Repeatable Read. MariaDB's documentation claims instant, lag-free replication with no lost transactions, but in practice the system behaved weaker than Read Uncommitted. Jepsen filed four bugs (MDEV-38974, MDEV-38976, MDEV-38977, MDEV-38999) and recommends users set innodb_flush_log_at_trx_commit=1, treat read-modify-write patterns as unsafe, and not rely on committed transactions being visible to subsequent transactions.
Sort: