Is it still recommended to use uvloop to improve performance?

[mcarans]

On the third party libraries page, it says for uvloop "We highly recommend to use this instead of standard asyncio." but uvloop isn't included in the speedups on the main aiohttp page.

I noticed a uvloop PR that says "Performance claims are not so valid in newer Python versions." There are also various open PRs particularly from user "tarasko" that suggest there are some performance issues.

Is it still recommended to use uvloop?

[Dreamsorcerer]

Probably not, though it would nice to see some benchmarks. We've also had bug reports caused by uvloop (the latest still hasn't been merged even though a fix is ready, so we've had to add workarounds here).

[tarasko]

I did a benchmark for various websocket clients against asyncio and uvloop. Hope this shed some light on real world performance advantages of uvloop (if there are any).

https://github.com/tarasko/websocket-benchmark

[Dreamsorcerer]

Seems a marginal improvement still. But, we can't recommend it until they merge the bug fixes, as it's been causing too many problems for users.

Would also be great to see similar benchmarks for server-side performance.

[tarasko]

Regarding server benchmark what do you think it should look like?
Should it be just one loop - one thread server that is bashed by multiple clients?
Would it be interesting to try freethreading python, multiple loop - multiple threads? Is it a real world use-case? I'm not sure if many libraries support freethreading yet?
Should clients send request, wait for response, send request again, or use some other strategy?

[Dreamsorcerer]

I guess our common use case is a single threaded server with multiple (or even a single) client. I think a lot of our optimisations were to help homeassistant deal with noisy websocket connections from IoT devices on the network, in which case it could just be 1 or 2 spammy clients sending to aiohttp on a RPi.

I'd test freethreading if it's minimal effort, but wouldn't expect much difference here. Not sure who's using it, but we had several nags to support it from users.

Should clients send request, wait for response, send request again, or use some other strategy?

I'm not sure. Maybe @bdraco has some insights based on previous optimisation efforts.