How does Uvicorn use asyncio to handle thousands of concurrent connections?

WHAT THIS TESTS: This question probes whether you understand how Python achieves concurrency without parallelism by default, and how an ASGI server leverages that model to handle high connection counts. Interviewers want to see that you know the event loop is not magical threading but explicit cooperative multitasking, and that you recognize the GIL boundary that forces multi-process architectures for CPU scaling.

A GOOD ANSWER COVERS: First, define the event loop as the core of asyncio, a single-threaded construct that uses selectors like epoll or kqueue to multiplex I/O and schedules coroutines. Second, explain that when FastAPI awaits an external call such as a database query or HTTP request, the coroutine yields control and the loop immediately switches to another ready task, allowing one process to interleave thousands of connections. Third, state that Uvicorn creates and runs this loop, accepting TCP connections and translating them into ASGI event dictionaries that it pushes into the FastAPI application. Fourth, emphasize that because of the GIL, only one thread of Python bytecode executes at a time per process, so a single event loop cannot utilize multiple CPU cores for computation; therefore Uvicorn relies on multiple worker processes, typically four to eight per machine, to achieve true parallelism across cores.

COMMON WRONG ANSWERS: A major red flag is claiming that asyncio spawns a thread per connection or that it removes the GIL. Another is describing the event loop as a separate thread that runs in the background; the loop runs in the main thread and callbacks execute there. Candidates also err by saying Python coroutines run in parallel on multiple cores within one process, which is impossible for standard CPython bytecode. Finally, confusing sync WSGI threading models with ASGI async models suggests a shallow understanding of the stack.

LIKELY FOLLOW-UPS: An interviewer might ask what happens if you call a blocking library like requests or SQLAlchemy sync inside an async path operation. They could also ask how you would profile event loop lag, or why you might choose Hypercorn over Uvicorn. Another common tangent is how to run CPU-bound work inside an async endpoint without stalling the loop, which leads to ProcessPoolExecutor or offloading to a task queue.

ONE CONCRETE EXAMPLE: Imagine a FastAPI endpoint that awaits an async PostgreSQL driver such as asyncpg to fetch a user. When the query hits the network, asyncpg yields to the event loop. Uvicorn, which owns that loop, immediately begins parsing headers from a new WebSocket connection. Once the database packet returns, the loop resumes the original endpoint coroutine and Uvicorn sends the ASGI http.response.start event back to the client. If the endpoint instead ran a synchronous bcrypt hash without using run_in_executor, the entire loop would freeze for fifty milliseconds and every concurrent connection would stall.

Source: fastapi.tiangolo.com