In terms of scaling, using a consumer/producer worker queue model is always a consideration. I like redis, personally and it can be used either with your own worker daemons distributed however or via one of the async frameworks you mentioned.

Your mentioning of SNMP made me look to see if observium had some sort of API and Perl client, but instead I found App::Netdisco - the author may have solved some or all of the problem you're addressing. I found it because it does have an observium plugin, App::NetdiscoX::Web::Plugin::Observium. Looking at the author profile on metacpan, I also saw: YATG (Fast SNMP data poller daemon, with storage and graphing) and YATG::Tutorial.