Every t-test and delta-method formula in this book assumes the split between arms is actually the fair random split it was configured to be. That assumption can be silently false — a routing bug, a caching layer, a client that drops some events — and the way to catch it is a chi-square test on the arm sizes themselves, run before trusting any effect size.
What Sample Ratio Mismatch Actually Is
Sample Ratio Mismatch (SRM) is what it's called when the arms of a randomized experiment come back a noticeably different size...