Linux Debian 6.0 .
Zabbix server and monitored hosts ar on different networks.

Description looks not finished.

And the idea is not very logical: network connection fails .... -> no alert is sent

In this scenario Zabbix is on a remote network and monitor hosts on different network.
If Zabbix network fails, we get hundred of alerts for all the monitored hosts which are working ok and the problem is zabbix server connection only.
Since the only way we get alerts is through network it make no sense to send alerts when network fails.
If zabbix server connection server fails we need to monitor it externally from a different location eventually using an othere zabbix server.
So it does not matter if you don't get alerts in case of local network failure, eventually in this case you can try to send only a single alert for the network connection failure that will be really sent when network is up again or using an sms gateway.
Normally we monitor zabbix server network connection by pinging google from zabbix server.

sounds like "easier way to manage dependencies"

Yes but I don't want to create manually a dependency for each trigger on each monitored host.
I forgot to tell that when network connection is up again zabbix should exit mainteniace mode.

ZBXNEXT-1375 may be related
See also http://zabbix.org/wiki/Docs/simplify_ad_hoc_maintenance