On Cloud Run, our backend answered its first request in 14 seconds. Almost all of that time went to one place nobody on the team would have guessed, and the profiler alone never found it.

This is the full hunt: the tooling that finally showed us the truth, the framework internals we had to read to understand what we were seeing, and the obvious optimizations that quietly made cold starts worse before anything got better. Image size, autoloading, the compiled container, cache warmup, and the way they interact.

We finished at 1.8 seconds. You'll leave with a method for diagnosing your own cold starts, not a list of tricks to copy.