Repository navigation
Description
As reported by a customer we're seeing the following behavior:
- App is launched
- A cold start is detected and an
ui.loadtransaction is created (app.start.cold+ttid) - The app is backgrounded before the first frame draws (and the activity only progresses through the
onCreate+onStartphases, noonDestroy) — so TTID never happens, and no lifecycle cleanup on our side runs. - The device enters Doze, which freezes our deadline timer so it doesn't fire at +30s.
- Hours later the device wakes up, the deadline finally fires when the app is resumed, and our
forceFinishstamps the spans with the current time. - The resulting transactions spans over multiple hours.
I would argue that we should drop any foreground launches which have a doze in-between, as it just skews the real "user perceived" launches.
On the other hand, a real user was probably waiting for a few seconds for the app to come alive, but it didn't, so the user bailed.
Related issue: #5530
Decision (2026-11-12)
Filtering on span.status:ok doesn't remove these transactions. After the device wakes from Doze, or the process is thawed by the cached apps freezer, the overdue deadline timer races the first frame:
- Timer first: everything ends
deadline_exceeded, but at the wake time, so it's hours long. - First frame first: TTID and app start are measured from the Activity's start, and the idle timeout finishes the transaction as
ok.
In one customer's data, most ui.load transactions over 60s were ok with this signature.
The fix goes on main and changes no timings: every timestamp, duration and measurement stays as it is. It measures the deadline with a Deadline on the boottime ticker, which counts deep sleep. Once that has passed, the transaction and every span that finishes after it get deadline_exceeded, the status the deadline timer already gives. So span.status:ok filters these transactions out, whichever path finishes them. The timer stays, so an awake device still ends on time. #5755 (closed) tried dropping late transactions instead.
Side effect: the extended app start vital is skipped when its span ends after the deadline, because AppStartExtension ignores deadline_exceeded spans.
Activity
linear-code commentedon Jul 10, 2026
JAVA-642
revert: fix(tracing): End timed-out transactions when the timeout fel…
runningcode commentedon Nov 6, 2026
We decided aren't going to make changes to
sentry-javato address this but rather we will advise customers to filter spans byspan.status:ok. When the deadline timer fires, the spans are marked asSpanStatus.DEADLINE_EXCEEDED.github-actions commentedon Nov 17, 2026
A PR closing this issue has just been released 🚀
This issue was referenced by PR #6599, which was included in the 8.65.0 release.