Hate to be the Monday morning quarterback but hey, it's Saturday morning and I'm stuck in a Starbucks in Benicia. So, this is a nice walk through the optimization process but it is a fundamentally unscalable system. When the workload triples next year, does it make sense to scale up the hardware? (No)

This system doesn't scale out without ad-hoc partitioning. I think Etsy's approach here should have been informed by the New York Times' project a few years ago. They converted 4 TB of scanned tiffs (their article archive) into PDFs on a hadoop cluster running on ec2 and s3. Parallelizing the process across hardware nodes is the scale free way to do this, exactly what hadoop was intended for.

All that said, I know there are smart folks at Etsy. I'll give 'em the benefit of the doubt; there may be a good reason not to go that route but this write up didn't make that clear.