Seeing what looks like a memory leak

[cpoppema]

For two projects I replaced uvicorn with uvloop with granian and my monitoring tools show an increase in memory usage. I am wondering if this is something I have to accept with granian - I don't know if it'll flatten out at some point - or if there is an underlying issue I might be able to help debug. Benchmarking is all about request numbers but does not mention any memory increase, so I suspect an issue.

Figures below are from 2 services running fastapi + pydantic + sqlalchemy.

service 1, graph is from 7 days:

cpu before (uvicorn)/after (granian)

requests before (uvicorn)/after (granian)

memory before (uvicorn)/after (granian)

service 2, graph is from 2 days:

cpu before (uvicorn)/after (granian)

requests before (uvicorn)/after (granian)

memory before (uvicorn)/after (granian)

[gi0baro]

My best guess at the moment is that might be related to mimalloc allocator.
Is that Linux on an x86_64 architecture?

[cpoppema]

This morning I overlooked the official 1.2.3 release, so I will redeploy with your release to see if it makes a difference. 🤞 The memory from my manual build so far looks like this:

[gi0baro]

@cpoppema thank you for the updates.

I doubt official 1.2.3 will make any difference, it seems there's definitely something going on here.
Probably the code is missing some Python GC calls (https://pyo3.rs/main/class/protocols#garbage-collector-integration), but given #265 is changing a lot of things, I need to schedule memory allocation tests after that is merged, and thus this will probably see some enhancements with 1.3.x series, which I expect to release in the next couple of weeks.
Thank you for your patience.

[cpoppema]

No problem and you're right, it makes no difference (orange=version 1.2.3 from pypi).

Feel free to @ me if want me to test a new version. (In the meantime I will most likely switch to hypercorn as a replacement for uvicorn)

[gi0baro]

@cpoppema can you provide additional details on how do you run Granian on such services? eg: number of workers, threads, blocking threads, threading mode..

[cpoppema]

I have been running with the following command:

granian --interface asgi service.server:app --host=0.0.0.0 --loop uvloop