On the one hand, this is really bad. A database that can't keep its consistency promises can make for a truly bad day.
On the other hand, the consistency promises Galera claims already aren't enough for critical applications even if true (serializable or GTFO), so I would hope that no serious user was running it in a way that has these problems.
Specific guidance if you need to use Galera today:
- Do not run Galera as a way to scale transactional write performance. It will eat your data.
- Running Galera as a way to scale read performance is probably fine, as long as all writes and all critical reads go to a single node.
- Running Galera as "Master-Slave-Slave with automatic failover" is probably best.
In the acceptable use cases, make sure that your chosen load balancer strongly enforces the single node requirement. One easily overlooked point is that you can't simply chose a different node for new connections on failover - if the LB chooses to switch, it needs to cut all existing connections, too.
It should go without saying that your application will need to handle loss of connection safely (because that can always happen for any reason), but…it probably doesn't. Go yell at your developers.