MariaDB Galera Cluster Conflicts and Certification Failures

How to recognise, diagnose, and recover from Galera-specific failure modes : certification deadlocks at COMMIT, hot-row contention across nodes, split-brain and Non-Primary partitions, gcache exhaustion forcing a full SST. Galera is a synchronous multi-master cluster built on certification-based replication ; its failure surface is different from standalone InnoDB and must be reasoned about as a cluster property, not a per-node property.

Quick Reference

• Galera ER_LOCK_DEADLOCK (error 1213) returned at COMMIT is NORMAL behavior, NOT a bug. The certification protocol detects write-write conflicts between nodes only at commit-time. ALWAYS retry the entire transaction.
• wsrep_local_cert_failures is a monotonic counter of certification failures on THIS node. Growing fast = hot-row contention across nodes. Redesign with sharded keys.
• wsrep_local_bf_aborts counts local transactions aborted by an incoming applier (brute-force abort). High value alongside cert-failures confirms multi-master write conflict pressure.
• ALWAYS retry with exponential backoff (50 ms base, 2x growth, max 5 attempts, optional jitter) to avoid retry-storms that compound the conflict.
• A 2-node Galera cluster is UNSAFE : it cannot form a stable quorum on a partition. ALWAYS run 3 nodes minimum, OR 2 full nodes + 1 garbd arbitrator.
• pc.weight (provider option, default 1, dynamic) biases quorum on asymmetric DC topology. Set the larger DC higher so it survives a WAN split.
• pc.ignore_sb=ON is DANGEROUS in multi-master : it allows writes on a disconnected partition. Use ONLY on a read-only standby or single-master topology, NEVER as a "quick fix" for split-brain alarms.
• wsrep_cluster_status='NON_PRIMARY' means this node lost quorum and is read-only. Find the minority partition and recover it ; do NOT force pc.bootstrap blindly.
• NEVER disable wsrep_on to "make the error go away". It de-syncs the node from the cluster and corrupts cluster consistency.
• Galera replicates ONLY write-sets (modified rows). Long SELECT-only transactions on hot data do NOT cause certification failures on the executing node.

Galera Failure-Mode Identification