UI load duration is hours long

[serglom21]

Integration

sentry-android

Build System

Gradle

AGP Version

N/A

Proguard

Enabled

Version

7.20.1

Steps to Reproduce

When looking at some of the UI load traces for an Android app, I see traces with hour-long durations, sometimes even going up to days. I see that traces with long durations are reported to have happened while the app was in the foreground, which removes the possibility of them being recorded while the app was in the background (that or the SDK is recording in_foreground wrongly).
Looking at the related issues, there are a lot of ANRs happening on that transaction. It seems that idleTimeout might not have any effect here since there is a span ui.load_initial_display which is active throughout the full trace timeline.

Expected Result

I would expect UI Load traces to not be hours / days long.

Actual Result

Traces are hours long and there are a lot of repeating onStart / onCreate lifecycle spans, which seems odd.

There is sensitive information on these traces so I am attaching them in the customer case

[philprime]

Hi @serglom21, thanks for creating this issue.

@markushi @stefanosiano could this be related to recent changes in cold/warm/hot startups?

[stefanosiano]

this comes from an internal discussion. Pasting here the last messages as a reference for discussion

Roman
yeah that doesn't look right, could this PR be the cause ?
Repeating onCreate/onStart spans do not look right for a single activity tx

Stefano
not sure what the problem could be
in a single activity the transaction is finished after the first frame, and even if we are repeating onCreate/onStart spans, only the last one would be attached and sent per transaction

There are a lot of ANRs happening on those trace events - since the main thread is frozen, could they prevent the SDK from ending properly?

nope. Either the main thread is frozen and the app is then killed (and our SDK is finished in this case) or the main thread gets “unfrozen” and continue working again (together with our SDK)

Could the SDK be tracking ui.load operations across different instances and accumulating these durations?

It shouldn’t. There was a fix in v 7.22.1 to ensure it doesn’t happen. I think it’s worth asking them to upgrade to this version and check if it fixes it for them