Design-wise: Two queue worker processes independent of web server processes. Delayed::Job lets you assign tasks a priority and lets you assign queue workers to priority levels it is allowed to look at. By convention, priority 0 (highest) is interactive tasks in my applications (a user is at their keyboard waiting for an answer) and priority 10 (least priority) is, well, Mixpanel. Not because I don't love them, just because if Mixpanel blocks for an entire month my rent still gets paid and that isn't true of any other priority level.
Queue worker A only works on priorities 0 through 9. Queue worker B works on 0 through 10. This ensures that even if Mixpanel (or Kissmetrics, also on 10) perpetually times out the higher-priority levels will never be totally blocked.
Delayed::Job worker processes are monitored by god, which resets them if they fail, become bloated, etc. They're also separately monitored by Scout monitoring's DJ plugin, which fires a yellow alert if a job is ever older than an hour (shouldn't happen but can in event of e.g. an API outage) or a red alert if there's ever more than X jobs in the queue (high probability that this represents a crash god can't recover from).
Because even that setup was letting through about one outage every six months to a year, I have one other ace-in-the-hole: my ajax polling actions which check whether particular jobs are complete will, if they fail Y times in a row, a) instantiate a queue worker within the web server process (this degrades request processing but can't totally break the site) and b) fire off an independent-from-everything-else phone call to my cell phone saying "Your first, second, and third line of defense have failed. Get ye to an SSH terminal."
I'm not the GP but in my startup we just use Resque, it works wonderfully. I've also had semi-good experiences with RabbitMQ via AMQP. There's nothing special about the structure. Depending on the system you'll use channels/queue names/routing keys/whatever when enqueuing a job and then workers will pull the jobs they're in charge off out of the queue. In the case of mixpanel jobs you should set it up so they stick around when they fail
