I have recently tried to convince the Python community to add a feature to the Python GC which would accept hints that a certain group of objects is unreachable and can be deallocated. Such a feature would reduce the amount of time spent in full memory scan and improve P99 latency. The actual hints can be provided by developers or Claude Code.
See Improving Python Garbage Collection Performance by Providing Unreachable Cycles Hints - Ideas - Discussions on Python.org for more details.
I need to test this first on a large scale complex Python environment, and TensorFlow seems perfect for this. Can anyone from the TensorFlow team with me on this? I can code the feature, but I will need a real-life environment to see the actual performance gain. If we can show definite improvement we can use it to push this feature to the Python team.