I've had git bisect work perfectly to show me that a commit that I would swear up and down couldn't cause the problem was what was causing the problem. In those kinds of situation git blame and git log are a bit useless since you can be looking right at the commit in question -- even looking at the diff in question -- and fail to actually 'see' it and go chasing down other rabbit holes that seem 'obviously' much more likely to be the cause. And this was typically after some customer updated for the first time in several years and so there was several years of changes to sift through to find the regression.