Message6148

Fixed in r7135

The actual problem is not in socket support, but in threads. In my previous change here, I was wrong about the Guava map in use, it's actually the canonical map org.python.modules._threading._jthread_to_pythread. The solution is to have it both weak keys AND weak values, similar to what we saw with class_to_type.

I can now run Gareth's example code in excess of 200,000 times (without the time.sleep of course, I have no patience :) for that), which is the point that the usual heap allocation of 512M would be exhausted on my laptop.

So we need to consider how to test this in regrtest. Probably the best bet is to use the SocketServer code in a subprocess with a minimal heap, run it for the equivalent of 10s or so, and verify that it exits without error.