Why does mascale sometimes fail to authorize a user despite valid credentials?

Question

linux

Hello.
There is maxscale + Galera Cluster (2 nodes Percona 5.7+ arbitrator), configured to separate write operations (we write to only one node) and reads (read from both nodes).
The problem is that in the maxscale logs once a minute or a little less often there are entries like

Sep 14 21:03:01 maxscale2 maxscale[9394]: [MySQLAuth] Splitter Service: login attempt for user 'username'@[10.20.30.40]:54198, authentication failed.
Sep 14 21:03:01 maxscale2 maxscale[9394]: [MySQLAuth] [Splitter Service] Loaded 55 MySQL users for listener Splitter Listener.

At the same time, in applications, an error about unsuccessful user authorization, like

php: /var/lib/asterisk/agi-bin/lib/adodb/drivers/adodb-mysqli.inc.php[109]: mysqli_real_connect(): (28000/1045): Access denied for user 'username'@'xxxx ' (using password: YES) to database 'DATABASE'

At the same time, a huge number of requests from the same applications with the same credentials are processed normally.
The set of accounts on both nodes is identical, the rights and passwords are identical. At the same time, about 600 connections from one application (one account) and about 50 from another (another account) are working - errors are observed here and there.
A maxscale machine does not lack CPU, memory, or disk, and the load is small.
On the galera nodes, there are no messages about failed authorizations in the logs (because the decision to authorize / not authorize is made by maxscale itself based on the contents of mysql.user), but there are messages like

2017-09-15T08:43:18.376924+03:00 1038984 [Note] Aborted connection 1038984 to db: 'DATABASE' user: 'username' host: 'xxxx' (Got an error reading communication packets)

Where to dig, comrades?

Answer 1

Updated from 2.1.3 to 2.1.7 - I've been testing for five hours, the problem has not yet appeared.