How would you diagnose memory growth in a Python service?

Production capture must fit the incident budget because deep allocation tracing adds overhead. I compare affected and healthy workers on the same build and traffic shape, and include container limits and OOM termination history. Native-heavy libraries may need their own profiler or allocator evidence. After finding ownership, I fix the retention path, add a soak regression and memory budget, then verify RSS stabilizes across several worker lifecycles rather than declaring success after one garbage collection.