How often should we measure performance? We could do so only when we make changes in the critical path, but [what if seemingly unrelated changes affect translation unit boundaries, and your compiler optimizations are affected?]\  What if you are consistently making changes to the critical path because you’re constantly trying to deliver better performance for your users? Getting into the business of predicting which changes will affect performance can be dangerous. In most cases, _you should be regularly testing for performance regressions. _ With regular commits from over 100 contributors, we’ve opted to measure this nightly, and for changes we expect to directly impact performance, we additionally measure and evaluate before committing.

#databases #p-values