I'm not certain what vps each server is outputting but I get the feeling it's in the range of 50-100/s
Graph shows with / without Persisted Buffer. Normally this data is on a 10s interval.
systemctl stop zabbix-agent2 also times out when persisted buffer is enabled. Likely the message queue never drains.
strace seems to indicate that there is alot of file open / closing and that every event looks like it's being written then read from disk. Since the persisted buffer is only needed when the remote server is unavailable surely the memory cache should be first used.
Also bulk inserts and prepared queries would help here for performance. I don't think our data rate is that high that this should be an issue.