I just tested using this setting, but it made no difference:
innodb_max_dirty_pages_pct_lwm=0.001
Tested with 2 existing nodes running and sync'ed, and one new node being attached.
As soon as SST sync ended, the first 2 nodes started piling queries, mostly stuck in commit.
If the new node was used, all queries got stuck in opening tables.
Please see attached screenshots showing part of the queues, which quickly reach 100+.
It took about 20 minutes to recover from that situation without using the new node at all! It took another 20 minutes for the new node to become usable (eg not too slow).
Restarted server, it did an IST, exact same result, just worse, pending queries above 200 this time!
How is it possible for queries to be stuck for several minutes (10!!!) while in nominal state they take less than 0.2 seconds !?
Saw this in logs just after the IST, however table_open_cache_instances = 16 is set in config file and live value is indeed 1!?
2022-07-27 23:55:09 11 [Warning] Detected table cache mutex contention at instance 1: 30% waits. Additional table cache instance cannot be activated: consider raising table_open_cache_instances. Number of active instances: 1.
FWIW we have 350+ DB each with 100 tables, and doing a mariabackup requires ulimit -n 919200, servers are processing about 3.000 queries/seconds during those tests, while it peaks at 10.000 during the day. A single server is capable of handling the peak load (verified earlier this morning) !