Comment 56 for bug 1596075

The slowness is definitely a result of lots of retries. I think we can look at that as a separate bug though. First I want to get it where everything is correct then worry about performance.

You should get a significant improvement just from using a single-writer deployment for Galera. Galera manifests failures as latent deadlocks on commits when using multiple writers so they are very expensive to recover from. With a single writer, constraint violations and staledata errors are recognized locally before it comes time to commit.