At LayerVault, we're using delayed_job (https://github.com/collectiveidea/delayed_job) for our queuing system. Everything goes through a database, so the performance won't be as good as something in memory. Also, I feel safer with the workers only communicating with queue servers and not with application servers. Just a litte more abstraction :)