> I'm afraid nailing anything here is hard.  For example, I don't know what you did to "measure memory", but if you're using stats reported by the OS, that's fraught with dangers too. 

I am interposing a custom allocator to track all the used blocks in all used pools in all arenas and then monitor all the memmap/malloc allocations. I get the same results if I don't use obmalloc (setting the PYTHONMALLOC=malloc) env var, which is more straightforward as I can just interpose a malloc implementation using LD_PRELOAD or a simple custom allocator. I don't claim to be exact but is better than just checking the resident size.

I didn't want to use sys._debugmallocstats() so printing to stdout does not impact performance, as I was also checking that the code does not make the interpreter slower.