Very interesting. I wonder if they tried running MongoDB on SSD. I would think that might mitigate having to have their working set in memory and potentially make the cold start problem a bit better. It'd certainly seem easier than migrating datastores.
If anybody would like a copy of the slides (sans video), I've made them available here: <a href="http://rob.by/2012/migrating-to-riak-at-shareaholic" rel="nofollow">http://rob.by/2012/migrating-to-riak-at-shareaholic</a>
Overall a great talk by @freerobby. One little nit... The Riak search feature is not built on lucene but modeled after it in terms of api. The search itself is not as robust as lucene but useful for most applications.