Use Python asyncio for Waiting, Not for Heavy Computation

A practical approach

Step 1

Confirm the work is I/O-bound: most time is spent waiting on networks, disks or databases, not computing.

Step 2

Use async-capable libraries throughout, such as an async HTTP client and an async database driver. Calling a blocking library inside a coroutine blocks the whole loop.

Step 3

Start concurrent work with asyncio.gather or, on Python 3.11 and later, asyncio.TaskGroup, which also cancels sibling tasks if one fails.

Step 4

Limit concurrency with asyncio.Semaphore so you do not open hundreds of connections at once or exceed an API's rate limit.

Step 5

Move unavoidable blocking or CPU-heavy calls off the loop with asyncio.to_thread, or into a process pool for real CPU work.

Worked example

A script checks 500 course links by requesting each URL in sequence and takes about eight minutes. Rewriting it with an async HTTP client, a semaphore limiting concurrency to 20 and a TaskGroup brings the run down to well under a minute. A later version that also parses large PDFs inside the same coroutines slows down again, because parsing is CPU-bound; moving parsing to a process pool restores the speed.

A mistake to avoid

Do not call time.sleep() or a synchronous HTTP library inside async code. They block the event loop and silently remove the benefit of async. Use await asyncio.sleep() and async clients instead.

Try it yourself

Write a script that fetches ten pages sequentially, then rewrite it with asyncio and a semaphore of three. Time both versions, then add a deliberately CPU-heavy function inside one coroutine and observe its effect on the others.

Your implementation checklist

• I use asyncio only for I/O-bound work
• Every library called in coroutines is async-capable
• Concurrency is limited with a semaphore
• Task groups or gather manage concurrent tasks and failures
• Blocking and CPU-heavy work runs outside the event loop