How would you profile a slow Python service without guessing?

During a live incident I start with low-risk telemetry and sampling, obtain approval for more invasive capture, and preserve the exact build and load shape. CPU throttling, garbage collection, database waits, and downstream timeouts can imitate application hot spots, so host and dependency evidence stays aligned with the profile. Success means user-facing tail latency or capacity improves without moving saturation elsewhere. The resulting regression test protects the measured path, not a synthetic microbenchmark alone.