How Can I Run Python Code in a Container That Respects cgroup Limits?

I don’t have a direct answer, but I've used Slurm clusters that will terminate Python processes if they exceed their memory limits, which might be worth looking into for managing Python workloads effectively.

It may help if your application developers implement custom environment variables to manage configurations better. For example, a thread pool could be limited through env vars, but they'd need to code that functionality. If you can't do that, just let the CPU throttling work as it does.

Python's memory management is definitely more rigid compared to languages like Go. Unlike those garbage-collected languages that can allow extra memory for performance gains, Python usually uses just what it needs. This makes it tough to enforce memory limits in the same way. Most Python objects are freed almost immediately rather than waiting for garbage collection, so trying to limit memory might not hurt performance that much. If you’re using threading or task schedulers, consider managing those at the library level since Python normally runs single-threaded unless specified otherwise. Also, remember that different Python interpreters could behave differently—like PyPy, which supports some garbage collection settings.

In long-running Python applications, you may experience memory fragmentation. As your app allocates and releases memory, gaps can form. For example, if a 100MB allocation is freed but smaller allocations fill in the gaps, the next time you need 100MB, you'll have to allocate it afresh. This can create a scenario that resembles a memory leak. A common fix is to restart the process regularly or use multiprocessing so that the OS can reclaim memory from subprocesses after they finish.

While Go has built-in features to limit CPU usage effectively in containers, Python applications typically handle threads and workers separately. You may want to cap the number of workers via environment variables or other deployment settings. If memory consumption is a problem, think about running Python apps in their own Docker containers to isolate memory issues from other apps.