A (long) while ago, one of my coworkers spent months putting together an enormous framework for processing some data. It was really well done - simple API, good error messages, verification at various steps, and reliable results. It was a little slow though. It was pushing 1 hour to complete a cycle, which was bumping up against other processes. A senior guy had a look at it, noticed that we could encode the data as a vector and apply some linear transformations to it, then decode afterwards. Thanks to lucky cache hits and BLAS, it took about 25 seconds to run.