In my last paid web development job, I had responsibility for a system that grew over time into a tangled mess. Almost every page view spewed dozens of warnings to the error logs. No one ever looked at them.

In the course of my refactorings (and writing tests), I silenced all of those warnings. Thus if we ever saw a warning in the error logs, we knew something unexpected had happened -- such as a customer doing something unanticipated.

This allowed us to refine our tests based on what our customers actually did, not what we thought they might do.

I would not trade that away for some perceived benefit of several milliseconds -- especially in an application where the greatest performance drain came from network latency.