Assess and improve session replay performance using camera view finder

[philprime]

Description

• We need to assess how session replay recording affects video capturing previews, especially using AVFoundation and AVCaptureVideoPreviewLayer
• Masking preview or disabling recording of preview could be complicated, as it is implemented as a CALayer, rather than UIView

[kahest]

We have one customer reporting performance issue with camera view finder in their app using SDK 8.41.0. They didn't yet update to 8.43.0, which is the first SDK release where SR is GA. So we should try to reproduce with both/current SDK version.

[kahest]

Also to test whether the performance improvement shipped with the experimental view renderer in 8.47.0 affects this issue

[philprime]

I created a sample app in #5055 to verify the behaviour with an orange background, a label "Background" and a preview view on top of the label. This is what the screen looked like:

With enableFastViewRendering = false this is the replay:

https://sentry-sdks.sentry.io/explore/replays/28d11e2f3f394a57a46aca12fb3937b6/

With enableFastViewRendering = true this is the replay:

https://sentry-sdks.sentry.io/explore/replays/365ced0a9ff34025a136834e43c48caa/

It seems like we are not able to render the preview layer using either of the render methods.
According to this StackOverflow post it might not even be possible to render the layer due to the inner workings of CPU vs GPU render graphics.

Regarding the performance:

At least on an iPhone 14 Pro I could not see any difference in performance between using the experimental view renderer or not. We might have to perform additional testing on older devices and more complex view hierarchies.

EDIT:
@kahest I believe we should see the AVFoundation camera input as a known limitation. If there is higher demand we could look into custom drawing logic.

[kahest]

At least on an iPhone 14 Pro I could not see any difference in performance between using the experimental view renderer or not. We might have to perform additional testing on older devices and more complex view hierarchies.

To increase confidence I'd like to run some simple tests on an iPhone 8, like we did for #4940.

I believe we should see the AVFoundation camera input as a known limitation. If there is higher demand we could look into custom drawing logic.

That's fair, we should add a note on the docs.

[philprime]

I did additional testing with an iPhone 8 using the sample app that will be merged with #5055.

Visually there is no difference noticeable, but Xcode instruments show that the new experimental view renderer has actually more app hangs than without:

Experimental View Renderer:

Default View Renderer:

We might be able to do a more thorough performance analysis with a real use-case, rather than the minimalist sample one.

EDIT: After reviewing the performance tests again it is now visible that there is a clear performance boost reducing view render time from ~157ms to ~40ms