A lightning-fast asyncio server for Python 3.
Install using pip:
pip install uvicorn
app.py:
def hello_world(message):
content = b'<html><h1>Hello, world</h1></html>'
response = {
'status': 200,
'headers': [
[b'content-type', b'text/html'],
],
'content': content
}
message['reply_channel'].send(response)Run the server:
uvicorn app:hello_worldimport asyncio
async hello_world(message):
await asyncio.sleep(1)
content = b'<html><h1>Hello, world</h1></html>'
response = {
'status': 200,
'headers': [
[b'content-type', b'text/html'],
],
'content': content
}
message['reply_channel'].send(response)Run the server:
uvicorn app:hello_worldThe server is implemented as a Gunicorn worker class that interfaces with an ASGI Consumer callable, rather than a WSGI callable.
We use a couple of packages from MagicStack in order to achieve an extremely high-throughput and low-latency implementation:
uvloopas the event loop policy.httptoolsas the HTTP request parser.
You can use uvicorn to interface with either a traditional syncronous application codebase, or an asyncio application codebase.
These are the same packages used by the Sanic web framework.
- I've modified the ASGI consumer contract slightly, to allow coroutine functions. This provides a nicer interface for asyncio implementations. It's not strictly necessary to make this change as it's possible to instead have the application be responsible for adding a new task to the event loop.
- Streaming responses are supported, using "Response Chunk" ASGI messages.
- Streaming requests are not currently supported.
Using wrk -d20s -t10 -c200 http://127.0.0.1:8080/ on a 2013 MacBook Air...
| Server | Requests/sec | Avg latency |
|---|---|---|
| Uvicorn | ~34,000 | ~6ms |
| Meinheld | ~16,000 | ~12ms |
This worker class interfaces directly with an ASGI Consumer.
This is in contrast to Django Channels, where server processes communicate with worker processes via an intermediary channel layer.