Imports, sys.modules, circularity

The import system: modules, caches, and main

import foo performs a search, then a cached load:

1. Check sys.modules — the module cache. Already imported? Reuse the SAME module object (imports are cached per-process).
2. Find a finder willing to handle the name (sys.meta_path); default finders search sys.path for packages (directories with __init__.py) and modules (.py files).
3. Execute the module body in a fresh namespace, cache it in sys.modules, and bind the name in your namespace.

Consequences that explain everyday mysteries:

• Import side effects run exactly once per process — the second import is a dict lookup.
• Circular imports break when module A (mid-execution) triggers import of B, which imports A back and gets the half-built namespace. Fixes: reorder imports, import inside functions, or restructure the dependency.
• __name__ == "__main__" is true only when the file is run as the entry script; when imported, __name__ is the module path — that is why the idiom guards CLI/demo code.
• .pyc bytecode caches (__pycache__) skip recompilation per source version — they are an optimization, invalidated by mtime/size.
• sys.path manipulation and namespace packages exist; prefer proper packaging (next module) over sys.path.append hacks.