[gregbayer]

Database is a pretty broad term these days. Doesn't this depend on the database we're talking about?

Something like Big Table is pretty good at storing small blobs. It's also pretty good at scaling, so you should never really have to delete "usage statistics, metrics, GPS locations, session data" or anything really.

Regarding logs, I completely agree. Something like tornado + scribe + s3 works quite well. (check out https://github.com/gregbayer/logcatcher) But then EMR + Scribe can be used to access it like a db.