Cluster hangs from highly concurrent sysbench session to multiple nodes

Hi,

For this exercise I used virtualbox, 3 nodes and maxscale on 1 node. I do not think maxscale has to do with it but it is an easy way to make it run on multiple nodes. I used the training VM's from https://downloads.mariadb.com/training/courses/mariadb-enterprise-cluster/OS-Images/.

Steps to reproduce:

• Install a cluster. I used a 3-node cluster
• Install maxscale using the configuration attached. I used a readconnroute router that only routes to the slave. This means that the traffic will be round-robinned between the nodes that do not have wsrep_local_index set to 1.
• Install sysbench
• Run sysbench to mascale

In most cases the cluster locked up after 10 seconds, in another case it took a couple of minutes. The sysbench command used:
sysbench --db-driver=mysql --test=/usr/share/sysbench/tests/include/oltp_legacy/oltp.lua --mysql-user=galera --mysql-password=galera --mysql-db=test --oltp-table-size=25000 --report-interval=5 --max-requests=0 --time=300 --num-threads=5 run --mysql-host=galera4 --mysql-port=4006

In the processlist of galera2 you will find a wsrep thread that is locked by something:
2 | system user | | NULL | Sleep | 3317 | Update_rows_log_event::find_row(129056) | UPDATE sbtest1 SET c='21925224570-87741851440-47847350341-42585377753-88509468277-36281519091-414733 | 0.000 |

On the other nodes you see threads waiting on COMMIT.

I am still waiting for the error logs, is there any other information you need?

Thanks,
Michaël