This is why you should benchmark like you test.

Spot regressions early (locally) but make the decisions based on the big the picture.