Hi,

Following on from my other thread, I'm testing MaxScale fronting a 2-node Galera cluster (with a Galera Arbitrator acting as the third node).

My current test outline is as follows:
  1. Identify the Write Master server in MaxScale
  2. Connect two mysql clients to MaxScale, and put them both into "READ COMMITTED" isolation level ("set global transaction isolation level read committed")
  3. Begin a transaction in each client session
  4. Take down the non-Write master node in the Galera cluster ("systemctl stop mysql" on that server)
  5. Commit the transaction in client 1
  6. Confirm that the clients can see the committed change
  7. Start up the stopped node in the Galera cluster ("systemctl start mysql" on that server)
  8. Commit the transaction in client 2
  9. Confirm that the clients can see the committed change
  10. ...
What happens is that everything appear to work up until step 8, when client 2 loses its connection to the database when trying to commit, reconnects and loses its uncommitted transaction as a result. At step 9, client 1 announces that it has lost the connection and reconnects.

Is this meant to happen?

If I run the same test, bypassing MaxScale and connecting both clients directly to the same Galera node, I don't get the connection loss.

I understand that MaxScale will route writes to one node and reads to another, but surely the loss of connection to a read node should not result in the loss of connection to the write node?

Am I doing something wrong?

Yours very puzzled.
PC