Fixed a fatal memory leak by disabling reload and show

[MicaelJarniac]

The application I'm developing had a fatal memory leak, where after a couple of hours of being deployed, even with no one using it ever, the memory would grow continuously until it got killed.

The fix? Simply disabling reload and show on ui.run:

if __name__ in {"__main__", "__mp_main__"}:
    ui.run(reload=False, show=False)

The before and after (notice the vertical line on the graphs).

This not only solved the memory leak, but also cut out idle CPU usage almost entirely.

My application was deployed inside a Docker container, but that's about all that I know of. The rest of deployment was handled by someone else.

I could try to find more information if necessary.

This is possibly related to #4521.

[MicaelJarniac]

I'm sharing this here in hopes that it might save someone else a lot of time trying to hunt down a similar leak.

Also, it may be worth mentioning something related to it on the deployment docs I think, unless this is way too niche and somehow only happened to me.

[evnchn]

Definitely same as #4521, if the memory growth heuristics and behavior are anything to stand by.

For the 4 combination of options, show and reload, their effect is documented here:

#4521 (comment)

Still have no idea why it happens though. And the chance vanished since the author at #4521 did not respond.

@MicaelJarniac I would be glad if you can help debug why reload=False, show=False leads to memory growth as I don't think me, Falko or Rodja has ever reproduced it successfully.

[rodja]

Indeed. We would be super grateful to see code which reproduces the error, @MicaelJarniac. Can you create a minimum reproducible example?

[MicaelJarniac]

And a list of runtime dependencies:

dependencies = [
    "authlib>=1.6.1",
    "google-ads>=28.0.0",
    "logfire[fastapi,httpx,system-metrics]>=4.2.0",
    "loguru>=0.7.3",
    "nicegui>=2.22.2",
    "nicegui-aggrid-enterprise>=0.1.5",
    "pydantic>=2.11.7",
    "pydantic-ai[logfire]>=0.6.2",
    "pydantic-graph>=0.6.2",
    "pydantic-settings>=2.10.1",
]

[MicaelJarniac]

I'm trying to figure out how to test a minimal version of this locally, to see if I can reproduce the issue.

[MicaelJarniac]

As for how the complete version is deployed, what I got so far is:

• Standard GitHub Actions deployment
• To AWS EKS
• With Istio sidecar

[MicaelJarniac]

Okay, I managed to reproduce it locally with the full project and only Docker, so deployment seems to have nothing to do with it.

[rodja]

Great! Hope you can remove code in the setup until you have something you can share with us. 🤞