ASF GitHub Bot logged work on BEAM-3776:

                Author: ASF GitHub Bot
            Created on: 13/Mar/18 22:28
            Start Date: 13/Mar/18 22:28
    Worklog Time Spent: 10m 
      Work Description: scwhittle commented on a change in pull request #4793: 
[BEAM-3776] Fix issue with merging late windows where a watermark hold could be 
added behind the input watermark.
URL: https://github.com/apache/beam/pull/4793#discussion_r174305062

 File path: 
 @@ -247,8 +249,26 @@ private Instant addGarbageCollectionHold(
    * Prefetch watermark holds in preparation for merging.
-  public void prefetchOnMerge(MergingStateAccessor<?, W> state) {
-    StateMerging.prefetchWatermarks(state, elementHoldTag);
+  public void prefetchOnMerge(MergingStateAccessor<?, W> context) {
 Review comment:
   I moved the merging code out of StateMerging.java so that I could reuse 
WatermarkHold.addElementHold instead of passing in the watermarks to 
StateMerging and duplicating the late detection logic.  To keep everything 
located together I moved the other uses of WatermarkHoldState from StateMerging 
to WatermarkHold.  AFAIK the StateMerging class is a leftover from lazy state 
merging and doesn't provide much benefit any more.

This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:

Issue Time Tracking

    Worklog Id:     (was: 80098)
    Time Spent: 1h 40m  (was: 1.5h)

> StateMerging.mergeWatermarks sets a late watermark hold for late merging 
> windows that depend only on the window
> ---------------------------------------------------------------------------------------------------------------
>                 Key: BEAM-3776
>                 URL: https://issues.apache.org/jira/browse/BEAM-3776
>             Project: Beam
>          Issue Type: Bug
>          Components: runner-core
>    Affects Versions: 2.1.0, 2.2.0, 2.3.0
>            Reporter: Sam Whittle
>            Assignee: Sam Whittle
>            Priority: Critical
>          Time Spent: 1h 40m
>  Remaining Estimate: 0h
> WatermarkHold.addElementHold and WatermarkHold.addGarbageCollectionHold take 
> to not add holds that would be before the input watermark.
> However WatermarkHold.onMerge calls StateMerging.mergeWatermarks which if the 
> window depends only on window, sets a hold for the end of the window 
> regardless of the input watermark.
> Thus if you have a WindowingStrategy such as:
> WindowingStrategy.of(Sessions.withGapDuration(gapDuration))
>  .withMode(AccumulationMode.DISCARDING_FIRED_PANES)
>  .withTrigger(
>  Repeatedly.forever(
>  AfterWatermark.pastEndOfWindow()
>  .withLateFirings(AfterPane.elementCountAtLeast(10))))
>  .withAllowedLateness(allowedLateness))
> and you merge windows that are late, you might end up holding the watermark 
> until the allowedLateness has passed.

This message was sent by Atlassian JIRA

Reply via email to