Threads, processes, and when async is wrong

Async has one job: let a process do other work while it waits. Two kinds of work defeat it. The first is a library that blocks, a synchronous SDK, a driver with no async version, a file read on a slow disk, where there is an await-shaped wait but no await to express it. The second is computation, parsing a forty-page PDF attachment, tokenising a corpus, computing embeddings locally, where nothing is waiting at all and the processor is the bottleneck.

For the first, a thread. Hand the blocking call to a worker thread and await the result; the loop stays free while the thread waits. For the second, a process. Hand the computation to a worker process on another core and await that; the loop stays free while the other core works. And since Python 3.14 there is a third option for the second case: the free-threaded build, where threads run Python code in parallel without the global interpreter lock, which changes the arithmetic for CPU-bound work.