Python Async Concurrency

Python Async and Concurrency

Select the model

• Use synchronous code when concurrency adds no material benefit.
• Use asyncio for many cooperating I/O operations whose libraries are async-aware.
• Use threads for blocking I/O that cannot be made async.
• Use processes or interpreters for CPU-bound work after measuring serialization and startup costs.

Asyncio defaults

• Prefer asyncio.TaskGroup for child-task lifetimes and grouped failure propagation.
• Bound fan-out with a semaphore, queue, or worker pool; never create unbounded tasks from untrusted input.
• Put time limits at meaningful operation boundaries with asyncio.timeout.
• Let cancellation propagate. Catch CancelledError only for necessary cleanup, then re-raise it.
• Clean up files, sessions, locks, tasks, and subprocesses with context managers and finally blocks.
• Do not call blocking functions on the event-loop thread; offload unavoidable calls explicitly.

Failure and verification

Define whether one failure cancels siblings, is collected, or is retried. Handle ExceptionGroup only at a boundary that can make that policy decision. Test success, partial failure, timeout, cancellation, bounded concurrency, and cleanup without relying on arbitrary sleeps.