Skipped frames increasing by exactly 1 every 30 s on an idle session — trying to understand why my report was rejected

DonF •⚡

New Member
I'm not here to pick a fight, and I'm not asking anyone to reopen anything. I'm trying to understand one thing: why a report backed by measurements was rejected, because that's the only part I couldn't get answered.

OBS 33.0.0, Linux, 60 fps, no stream, no recording, no replay buffer. "Skipped frames" increases by exactly 1 every 30.0 s, indefinitely, with a stable phase.

I spent about four hours measuring before reporting anything, with an AI assistant helping me structure the results — the measurements are mine. I eliminated, one by one and each time by measurement: an empty scene (same result), the cameras, the capture filters, the 24 browser sources tested individually in an isolated Chromium, the docks, the audio clock, the PipeWire capture, the GPU (no throttling, no Xid), the refresh rate, CPU affinity, the workqueues and vsync. Over 300 s: 10 increments, exactly 30.0 s apart, 10/18000 = 556 ppm. The counter maps to `video->lagged_frames`, which has a single writer, `libobs/obs-video.c:825`, reached only when `os_sleepto_ns()` returns false.

I filed that as GitHub issue #14165 with the code chain and the numbers, and asked which evidence would be accepted. The account was blocked before any technical exchange, the issue was closed in one line, and a reply from another account asking the same question was locked as spam 13 minutes later.

So my question is a simple one, and a one-line answer is all I need: is this behaviour by design — the counter reporting the video thread re-synchronising its clock — or is it something you would consider a defect? If it's by design, I'll write it in my own notes and stop there. I'm not looking for a winner here, just for the reason.
 
Adding what I missed, after going further.

Two commits by your own maintainers go through this code path, February 2022:

  • 4e5906d2b (jpark37), "libobs: Clamp video timing for safety": os_gettime_ns and os_sleepto_ns may not match up exactly. Touches libobs/obs-video.c and adds the clamped_diff block inside video_sleep(), still there in 33.0.0, on the very line that counts the skipped frame.
  • 4f15f1062 (jp9000), "libobs/util: Fix rounding error with os_sleepto_ns()": os_sleepto_ns can return false on a deadline not yet reached. Touches platform-windows.c only.
So OBS itself documents the defect, and the accuracy fix landed on Windows only. Linux, macOS and FreeBSD share platform-nix.c, which is still a relative nanosleep.

On my machine, 60 fps, no stream, empty scene: 10 skips in 300 s, every 30.0 s. One single writer, obs-video.c:825.

Intended, or was the nix path never covered by that 2022 work?
 
Hi,

In review, the account was banned because it submitted an AI generated report in violation of our AI policy.

It is not our responsibility to validate your LLM's output, so the question of "is this a valid report" is not something we will be taking the time to answer. In the future, please ensure that you read and understand a project's contribution policy before submitting to the project.

Thanks.
 
Hi Fenrir,

Thanks for the answer.

I did not break the policy: one bullet asks for text "human written", another says AI tools are "heavily discouraged", and the only one that mentions a ban covers "low-effort or incorrect submissions". Which one applies here?

The tests are mine: four hours, by hand. The problem is real, AI or not — the fix published the day after my report shows it. And a ban leaves me no future on this project: tell me what I have to fix.


Thanks.
 
Respectfully here, you did break the policy, by submitting text that was not human written. We discourage their use in regards to working on problems related to OBS and the OBS code base, and require that all submissions to us are human-written, regardless if AI was used in research or debugging. We prefer that AI isn't use anywhere in the process, but it's not really possible to police what people do on their own PCs.

What is in our purview is what we want to accept, parse, and triage as part of our project and our repository. That does not include AI-generated output. I have no comment if the issue was valid or not, but I skimmed through it and it's extremely verbose and doesn't seem to have an actual point. It's claiming that the label of "skipped" is incorrect? I'm not even sure why this is being reported like this, if that's the claim. If the testing and discovery was all done by a human, the results should have been provided by a human. We are not interested in what your AI thinks of the report; if we wanted to know what an LLM thinks we can ask it ourselves.

Cheers.
 
I read and checked every line of that report before posting it, and nothing went up without me validating it.

The report is not about the word "skipped". Its first claim is a number: on an idle session, no encoder running, renderSkippedFrames goes up by exactly one every 30.0 seconds, at a stable phase — and its cause is os_sleepto_ns() sleeping relatively, drifting by about 9 µs per frame. That function has not changed since 28.0.0, back in 2022: the code is the same today.

You say you have no comment on whether it was valid, then that it has no actual point. Both cannot hold.

And you say yourself that it is not really possible to police what people do on their own PCs. In practice, the only thing that can be observed is the admission — and I am the one who made it.

I understand your position on AI, and I will not argue it further here. But one thing stands: the code has moved since, on the mechanism that report describes. That was the only point I had in filing it.

And it is a pity I was blocked too: I would have been glad to keep hunting other Linux bugs, on an OS I have used for a very long time.

Thanks.
 
Back
Top