When would you reach for threads, processes, or asyncio?

Use asyncio for high volume I/O where you want thousands of concurrent waits in one thread, threads for I/O work that only has blocking libraries available, and processes for CPU bound work that needs real cores. The deciding questions are whether the work waits or computes, and whether the libraries you depend on offer an async interface.

My first question is always whether the work waits or whether it computes, because that answer picks the tool. Waiting on network or disk means asyncio if the client libraries support it, since an event loop will happily hold ten thousand open sockets on one core and the memory per task is tiny. If the only driver available is blocking, I use a thread pool, because a thread parked in a socket read releases the lock and costs almost nothing. Computing means processes, or dropping into a native library that releases the lock itself. On a service I worked on we had both in one request path: an async handler fanning out to six internal APIs with gather, and a run_in_executor call wrapping an old blocking SDK we could not replace. That combination is worth naming in an interview because most real systems are mixed. The mistake I watch for is someone putting a heavy pandas transform inside a coroutine, because the loop then stops serving everything else while it runs.