@hugollm I think I do understand what you mean I just wanted to point out that different implementations have different advantages -> I worked for a shared hoster (easyname.at) where we added different docker containers for different users (5.1, .... , 7) and we actually had generate a fpm config for every user so we can keep them inside of their container + paths to disallow hacking (shared www-data access), but even than if a customer had 4 different pages for his customers (2 wordpress, 2 drupal) he still had to reset the cache for all of them. we had other caching mechanisms in place but still you can see that the missing orthogonal design created some circumstances we had to consider.
one reason not to keep it in memory could be arbitrary size of the opcode and that it's a trade of between time complexity (computation time) vs space complexity (storage-space) this is not really correct since the instructions needed to load something from the harddisk are quite costly + the slow speed of hardisks (ms vs ns) but on the other hand the MRU-Cache implementation of the harddisk could speed it up a little ... but those are just guesses since I did not read up on the python implementation. if you do and you found out why, please do tell :)