One of the StarCluster plug-ins is for Condor which is supported in their latest AMIs. Perfect for me.
Docker looks promising, but there are other existing services stacked on EC2 that address the needs of science workloads. PiCloud does exactly the things you're asking for: http://www.picloud.com/platform/ . And the folks at Cycle Computing use Condor to manage the largest jobs ever run on EC2: http://www.cyclecomputing.com/ . I'm still working on my own stuff based on Groovy and Condor which I call Gondor, but it isn't at all ready for others to use. One thing I have found to be great is that there is a MacPorts portfile for Condor which works dandy. Just "sudo port install htcondor && sudo port load htcondor". http://research.cs.wisc.edu/htcondor/HTCondorWeek2013/presentations/SingerL_MacPorts.pdf . I don't yet see a nice single workflow that gets us to an integrated reproducible published result at the other end like Elsevier's Executable Paper http://www.elsevier.com/physical-sciences/computer-science/executable-papers, but I think we'll be there soon.
It's tolerated for the same reason "free" services such as those provided by Google, Yahoo!, Facebook, and most any other important social media company display ads and do data mining on visitor and subscriber behavior. The research being conducted with the data collected in MOOCs is one of the most socially valuable results possible since it leads directly to better education for the world. As Andrew Ng has stated plainly, his primary concern in participating in Coursera is delivering the best education possible to the world's poorest people. Coursera A/B tests most every aspect of the student learning experience and makes decisions based on what results in the best student learning outcome. Exactly what better system are you proposing?