If zabbix server gets LLD data from active proxy - it processes the data in the same time, TCP connection stays opened, proxy's data sender is waiting a reply from server.
All this time history is not synchronized during a few minutes and many nodata() triggers on server gets PROBLEMS.
For example 3 LLD rules send ~300KB json data in one (or in a few consecutive) connection(s). Proxy's data sender is 100% for 3-4 minutes, while usually the data sender is just 4% busy.
A 1 hour maintenance of DB server is performed (backup), but it does not not affect zabbix server in general meaning, except of:
server's proxy poller gets 1000 (=ZBX_MAX_HRECORDS) values (2MB data, multiple LLD rules data), and spends 7 seconds to process it. Without exiting from a loop, it ask more history again (ZBX_PROTO_VALUE_HISTORY_DATA) and again gets 1000 vales (10MB data this time) and processes it 30+ seconds (how much exactly - unknown) and so on.
As we know there is an internal queue for proxy pollers (info from Andris Zeila), so only one proxy poller will be working with the proxy.
Discovered issue: history sync will be delayed AND configuration synchronization of the proxy is not performed at all during the period!
To resolve all the issues - LLD data processing should be offloaded from current proxy<->server communication.