[discussion] Asyncio speedups about devpi

[fschulze]

Thanks for bringing this up as a discussion first.

Before adding any complexity like this, is there some real performance benefit for actual use cases?

[Agalin]

It's hard to provide performance increase values without some kind of a benchmark.

Uvloop is faster than the built-in implementation - that's a fact confirmed by multiple sources over the web, it's even the recommended by modern ASGI servers (e.g. uvicorn). It reaches performance similar to Go, faster than NodeJS.

But how large of an impact it may have on Devpi is hard to say. It heavily depends on load distribution. The more mirrored packages accessed the bigger the speedup should be. Definitely would be higher if the main API was ASGI but with pyramid it's impossible AFAIK.

I strongly believe using those options is way less complicated than implementing replicas or requests-only workers. In my case mirror syncs were the main source of bottleneck from what I can tell - simple local package downloads were handled by nginx directly.

[Agalin]

Oh. Didn't mention. Obviously, both uvloop and aiohttp speedups onlly help with parts of the stack. Bottleneck may be in other sections but hard to say without benchmarks.

[fschulze]

I'm fairly certain that the other parts of devpi-server are the bottleneck for mirror speed. The async stuff only happens every 30 min per project, when the cache expires. So first step would be profiling actual load.

[Agalin]

In my case it was a bursty load - at some point you got a CI pipeline or two making 5+ concurrent requests for various packages - many of them requiring a refresh.

Unfortunately I'm no longer able to profile it due to a migration to a different deployment structure.

[gaborbernat]

I'd make uvloop an opt-in option not the default. Because as #998 uvloop is also written in C and would slow down support for new Python versions.

[Agalin]

I like the approach used by both aiohttp and uvicorn - "use if installed, provide a single extra which installs them".