[Improvement] Prevent OperationRepo's 5 second wait from being skipped, and continuously pulling when empty #2033
Conversation
Create test to confirm that OperationRepo.processQueueForever keeps poll when it should suspend the thread while waiting for the next operation. My theory is this polling every 5 seconds means the app process may never fully sleep, which contributes to battery drain. We moved code to a new getNextOps() function simply to make it testable, no logic changes in this commit.
Created a new private Mocks class to encapsulate creating the depended
mocks into a one liner each test can use. Some tests override the
provided mock behavior by calling every {} again.
Refactored OperationRepo.processQueueForever to prevent it from continuously pulling when the queue is empty. My theory is this polling every 5 seconds means the app process may never fully sleep, which contributes to battery drain. This commit makes the "ensure processQueueForever suspends when queue is empty" pass.
jkasten2
force-pushed
the
fix/op-repo-fix-keeping-cpu-awake
branch
from
207c797
to
4a64d97
Compare
jkasten2
changed the title
jkasten2
changed the title
jkasten2
requested review from
nan-li,
emawby,
shepherd-l,
jennantilla and
brismithers
| // 2. Wait at least the time defined in opRepoExecutionInterval | ||
| // so operations can be grouped, unless one of them used | ||
| // flush=true (AKA force) | ||
| val startTime = _time.currentTimeMillis |
There was a problem hiding this comment.
Wondering if I missed something about this logic, technically startTime is not necessary since this variable is never used or updated?
There was a problem hiding this comment.
ah good point, I removed it in a --fixup commit just now. I think I refactored this a few times and forgot I didn't need startTime any more.
| } catch (e: Throwable) { | ||
| Logging.log(LogLevel.ERROR, "Error occurred with Operation work loop", e) |
There was a problem hiding this comment.
This try-catch is removed because it does not seem to have a specific purpose or thing it is meant to catch right?
There was a problem hiding this comment.
Ya, there isn't anything that I can see that could throw, and if it does we want to know instead of silently moving on. Also note I left the try-catch on executeOperations since it wasn't in scope of this PR.
Was there an outcome to this? |
We renamed waitForWake to waitForNewOperationAndExecutionInterval to describe that it is doing two things. Fixed the logic where enqueuing two or more operations would skip our executionInterval logic. Added test coverage for this logic, as well as a 2nd test to ensure flush can still skip waiting for it. Lastly cleaned up the force state in the main processQueueForever loop, we were able to fully encapsulate it into waitForNewOperationAndExecutionInterval. The other part of the logic is once we start processing the queue we want to do so until it's empty again, which ops != null is the only checked needed in processQueueForever to cover that.
We don't want to do any waiting when we call wake, so use trySend instead. Also added result checking, to ensure we throw instead of silently handling the error so we know if there is an issue at any time.
jkasten2
force-pushed
the
fix/op-repo-fix-keeping-cpu-awake
branch
from
60a4280
to
fd60e70
Compare
After some longer tests it seems inconclusive now. Don't think it worth the time to try other tests right now, there is going to be a lot of other improvements in the upcoming 5.1.7 release (such as retrying with backoff) that might be the root source of some reported battery drain. |
jkasten2
changed the title
jkasten2
changed the title
Description
One Line Summary
Prevent
OperationRepofrom continuously pulling when empty, which might be putting extra load on the device.Details
This also fixes a bug where batching wasn't waiting 5 seconds in most cases.
Motivation
The pulling might be causing some battery drain due to the app's process never fully sleeping or CPU continuing to be woken up. However the theory is this may only be noticable factor on some devices (probably more noticeable on watches running Android).
Scope
Only changes to the
OperationRepopolling logic.Testing
Unit testing
Added a new test to confirm it continuously pulls when the operations list is empty.
Manual testing
Test setup
Tested the battery drain by unplugging the device for a long time then obserserving how much power it takes to rechanged the device to full again. Recharge was measured with a USB multiple meter, getting the Wh it took to recharge the device until it stopped accepting power.
We test 3 scenarios, baseline to see how much power the device loses over time without OneSignal, and then two more comparing before after. Results are as follows:
Baseline (OneSignal app not installed on device)
Test 1 - 5.1.6
Test 2 - 5.1.6 (with OperationRepo fix)
Summary of testing
The data is inconclusive, unfortunately this method of testing is resulting in wildly different numbers between tests, even tests that last 24 hours. There stands to be some theatrical improvement still from this PR, but requires a different testing strategy and/or different devices to prove.
Affected code checklist
Checklist
Overview
Testing
Final pass
This change is