Stories

As part of a post on web concurrency in Rails, I’ve been running benchmarks to compare the Puma and Unicorn Ruby HTTP servers. Puma performs better than Unicorn in all tests that were either heavily IO-bound or that interleaved IO and CPU work. In the CPU-bound tests where Unicorn performed better than Puma, the gap was small enough that Puma can still be considered a very good choice.

Tracking authenticated user sessions can be implemented in Redis using setex with some serialised JSON. It works pretty well until you have to cope with millions, or even tens of millions of sessions where the memory usage and performance can suffer.

By using Redis data structures more effectively we can achieve a 70% reduction in memory usage, at the cost of both code and conceptual complexity. Is it worth it?

Split testing is a cornerstone of how we improve our products. While we usually run such tests for user-visible interface changes, this is an example of running a (successful!) multivariate test between CDNs.

The outcome is a 7% improvement in asset load times globally, translating into a +1% conversion on our site.