MySQL high availability in 2026 — Galera, InnoDB Cluster, or async replicas?

MySQL high availability splits into three real options in 2026: async replication with manual or orchestrated failover, Galera-based multi-primary (Percona XtraDB Cluster / MariaDB Galera), or MySQL Group Replication via InnoDB Cluster. Each makes a different trade-off between commit latency, write conflicts, and operational complexity.

The wrong answer is "we'll just set up CHANGE MASTER TO and hope." We see this everywhere. It doesn't survive a primary failure with any data integrity story you'd want to defend.

A working InnoDB Cluster bootstrap

-- on each node
SET PERSIST group_replication_local_address='node1:33061';
SET PERSIST group_replication_group_seeds='node1:33061,node2:33061,node3:33061';
INSTALL PLUGIN group_replication SONAME 'group_replication.so';
 
-- on one node
SET PERSIST group_replication_bootstrap_group=ON;
START GROUP_REPLICATION;
SET PERSIST group_replication_bootstrap_group=OFF;

MySQL Router in front, configured to follow the primary. Failover is sub-15-seconds.

The full write-up covers:

• Async replication: when it's actually enough, and the orchestrators (Orchestrator, MHA) that make it survive failure
• Galera trade-offs: synchronous certification, the write conflict tax, and the gcache sizing nobody plans for
• InnoDB Cluster: simpler than Galera, but with its own failover quirks
• The decision tree we run with customers (read/write ratio, write concurrency, geographic spread)
• Backup interaction: XtraBackup on Galera nodes vs InnoDB Cluster
• Application-side gotchas (auto-increment offsets, read-after-write consistency)

We deploy this for every managed MySQL customer who needs HA.