Writing on product work that we keep coming back to, gathered from across the internet.
Every title opens the original on the publisher’s own site. We point at this writing and credit the people who wrote it. We do not host it.
Stopping a test the moment it looks significant can inflate the false positive rate several times over, and Miller shows the arithmetic behind that. Short, precise, and it settles the argument about peeking at results for good.
Kohavi and Thomke on running experimentation at scale inside a large organisation, including how many tests fail and why that is the point rather than a problem. Helpful when you need to explain to leadership that a low win rate is a working programme.
How Netflix structures cells, allocation and analysis so any engineering team can run a test without building infrastructure first. Concrete detail on what a platform has to provide before self service experimentation is realistic.
Five categories of assumption and four ways to test them, with a clear line drawn between an assumption test and an experiment. Torres also answers the objection about time, arguing that an hour long test beats the perfect one you never run.