The feature relies on pressure signals, which are supported everywhere except for Linux? +haraken@ to confirm.

On Desktop, memory pressure signals are inconsistent, and generally not very useful right now. +sebmarchand, who's working on creating useful windows memory pressure signals. I believe no one is working on macOS signals.

OK, so let's first do the experiment on Android only.

Sergey, do you want to take this?

Well, I can give it a try, but it won't be my highest priority atm. Also, I don't have any experience with Finch experiments, where do I start to set up a new experiment?

I think the process is not that complicated on dev/canary - basically we can "just" go and enable the experiment, but we need to have metrics ready.

We have metrics for total renderer memory, so we can see improvement there. But do we have metrics that would tell us how many times the feature caused additional buffering when user went backwards in the playback?

BTW there seems to be an issue in how reliable memory pressure signals are on Android, see b/73310928. It looks like the issue could be fixed on Chrome side though (which experiment could only be done in M66+).

dalecurtis@, servolk@, do we have UMA metrics that tell us how often users go back in the playback? How hard it would be to add such metric?

Ideally we need something like "how often users go back and have to wait for data". With such metric we could trivially evaluate impact of MemoryPressureBasedSourceBufferGC feature.

AFAIK there's no such metric

Tokyo memory team is happy to help this.

keishi or tasak: Would you help servolk to add the UMA and start a Finch experiment?

haraken@ was your team still looking at relaunching this experiment?

Note that the memory pressure signals on desktop (Windows at least) are still pretty inefficient and inconsistent across platform (you can't really compare the MemoryPressureLevel values [1] between different platforms as they don't measure the same thing).

There's still no ETA for the better memory pressure signal on desktop (this is tracked on crbug.com/771478).


[1] https://cs.chromium.org/chromium/src/base/memory/memory_pressure_listener.h?l=50&rcl=d33d841ba9903ec8ad2a6874b181cfa2efe260ca 

dskiba left the team and no one is looking at this.

For now I think we can focus on enabling the feature on low-end Android (where memory pressure is (not good but) more reliable than desktops).

I don't think it's terrible to enable the feature on desktops because it would be better than having nothing, but that would be a separate discussion.

Dale: Would it be possible to find an eng in your team?

Yup I'll take it.

Dan said he had bandwidth to pick this up.