(It later turned out that way before I joined, some other people also called the big problem out, but were overruled by someone high up, and after that everyone kind of forgotten that there was a problem in the first place.)
2) At my last job, I wrote a tracing profiler for Common Lisp - or rather, a quick and crude pile of hacks resembling a tracing profiler, now available here: https://github.com/TeMPOraL/tracer. Up until then, we had a sense that our performance problems came from the database, but having a hard time pinning them down with regular profiling, we blamed them on the database driver (that we've written ourselves). When we started using the tracing profiler, we quickly discovered that the driver itself was perfectly fine - it's the amount, shape and structure of the queries we're doing that caused our performance to suffer death by a thousand cuts. That couple hours of detour into writing the profiler paid for itself in the following week with an order-of-magnitude performance increase, and the insights gained shaped the further refactoring and redesigning efforts.
(Some more details about the profiler here: https://old.reddit.com/r/Common_Lisp/comments/et0fx1/tempora.... I was supposed to blog about it, but couldn't find the time this week.)