When any of Zabbix Server subprocesses (poller, trapper, whatever) dies, Zabbix server master process terminates all other subprocesses and dies deliberately.

There are some serious drawbacks to this:
1) Zabbix Server restart on large installations causes severe CPU usage on networking devices due to lots of SNMP requests
2) Zabbix Server restart on large installations can take several minutes for pollers to make it catch up with all the data due to lots of hosts offline, not coping with the initial load, etc.

So why not to have restart procedure for died subprocess instead of killing the whole server?

Also, if subprocesses are restarted, this can be used as a last resort means to fight with transient memory leaks by 'reloading' Zabbix server, effectively gracefully killing all subprocesses and restarting them back while retaining caches and stuff.