__enter__ runs at the top of the with block; whatever it returns is bound to the as target. __exit__ always runs when the block ends, whether by falling off the bottom or by an exception. The three arguments (exc_type, exc_value, traceback) are all None on a clean exit and describe the exception otherwise. Returning None (or any falsy value) lets the exception propagate, which is what 99% of context managers want. The pattern is the right tool whenever 'do X before, undo X after' has to survive every exit path.

__exit__ is the only place a context manager can decide whether to swallow an in-flight exception. Returning a truthy value tells Python to treat the exception as handled, which is how contextlib.suppress is implemented. Use this power sparingly: silencing exceptions hides bugs, so always check exc_type and only swallow the one you mean to. Returning False (or None) for every other exception keeps the default 'propagate' behaviour for unexpected errors.

The classic use case is anything with paired open / close semantics: sockets, files, DB cursors, mutexes. The class form shines when the resource has methods you want to call inside the block (c.query(...)) and state worth tracking on the instance (c.closed). Even when the body raises, __exit__ still runs first, so cleanup is guaranteed without a hand-written try / finally. The open_count invariant going back to zero in both branches is the proof: the with block survives every exit path.