[
https://issues.apache.org/jira/browse/BEAM-8543?focusedWorklogId=442037&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-442037
]
ASF GitHub Bot logged work on BEAM-8543:
----------------------------------------
Author: ASF GitHub Bot
Created on: 05/Jun/20 21:35
Start Date: 05/Jun/20 21:35
Worklog Time Spent: 10m
Work Description: kennknowles commented on a change in pull request
#11924:
URL: https://github.com/apache/beam/pull/11924#discussion_r436173779
##########
File path:
sdks/java/core/src/test/java/org/apache/beam/sdk/transforms/ParDoTest.java
##########
@@ -3938,10 +3939,11 @@ public void testEventTimeTimerOrdering() throws
Exception {
ValidatesRunner.class,
UsesTimersInParDo.class,
UsesStatefulParDo.class,
+ UsesUnboundedPCollections.class,
UsesStrictTimerOrdering.class
})
public void testEventTimeTimerOrderingWithCreate() throws Exception {
- final int numTestElements = 100;
+ final int numTestElements = 5;
Review comment:
Why shrink it? Does the test get really slow? Is this going to be a perf
problem overall?
##########
File path:
runners/google-cloud-dataflow-java/worker/src/main/java/org/apache/beam/runners/dataflow/worker/StreamingModeExecutionContext.java
##########
@@ -577,12 +583,21 @@ public void flushState() {
WindmillTimerInternals.windmillTimerToTimerData(
WindmillNamespacePrefix.USER_NAMESPACE_PREFIX,
timer, windowCoder))
.iterator();
+
+ cachedFiredUserTimers.forEachRemaining(toBeFiredTimersOrdered::add);
+ }
+
+ Instant currentInputWatermark =
userTimerInternals.currentInputWatermarkTime();
+ if (userTimerInternals.hasTimerBefore(currentInputWatermark)) {
+ while (!toBeFiredTimersOrdered.isEmpty()) {
+ userTimerInternals.setTimer(toBeFiredTimersOrdered.poll());
+ }
}
Review comment:
Yea I don't actually understand what this block is for.
FWIW to do timer deletion/reset cheaply without building a bespoke data
structure just keep a map from id to firing time or tombstone. This way,
whenever a timer comes up in the prio queue you pull out the actual time for it
from the map. If it is actually set for another time, don't fire it. If it is
obsolete, don't fire it.
##########
File path:
runners/google-cloud-dataflow-java/worker/src/main/java/org/apache/beam/runners/dataflow/worker/StreamingModeExecutionContext.java
##########
@@ -577,12 +583,21 @@ public void flushState() {
WindmillTimerInternals.windmillTimerToTimerData(
WindmillNamespacePrefix.USER_NAMESPACE_PREFIX,
timer, windowCoder))
.iterator();
+
+ cachedFiredUserTimers.forEachRemaining(toBeFiredTimersOrdered::add);
Review comment:
Do we even need `cachedFiredUserTimers`? It seems obsolete if we
populate the priority queue. The name is also wrong - even before this PR it
wasn't a cache. It is a lazily initialized iterator. Instead, we should have a
lazily initialized priority queue (like you do) and just a flag to say whether
the incoming timers have been loaded yet.
##########
File path:
sdks/java/core/src/test/java/org/apache/beam/sdk/transforms/ParDoTest.java
##########
@@ -4040,7 +4043,8 @@ public void onTimer(
}
};
- PCollection<String> output =
pipeline.apply(transform).apply(ParDo.of(fn));
+ PCollection<String> output =
+
pipeline.apply(transform).setIsBoundedInternal(IsBounded.UNBOUNDED).apply(ParDo.of(fn));
Review comment:
Should not be calling `setIsBoundedInternal` here. Is this just to force
streaming mode? We need to just create a separate run of ValidatesRunner that
forces streaming mode.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
Issue Time Tracking
-------------------
Worklog Id: (was: 442037)
Time Spent: 40m (was: 0.5h)
> Dataflow streaming timers are not strictly time ordered when set earlier
> mid-bundle
> -----------------------------------------------------------------------------------
>
> Key: BEAM-8543
> URL: https://issues.apache.org/jira/browse/BEAM-8543
> Project: Beam
> Issue Type: Bug
> Components: runner-dataflow
> Affects Versions: 2.13.0
> Reporter: Jan Lukavský
> Assignee: Rehman Murad Ali
> Priority: P2
> Time Spent: 40m
> Remaining Estimate: 0h
>
> Let's suppose we have the following situation:
> - statful ParDo with two timers - timerA and timerB
> - timerA is set for window.maxTimestamp() + 1
> - timerB is set anywhere between <windowStart, windowEnd), let's denote that
> timerB.timestamp
> - input watermark moves to BoundedWindow.TIMESTAMP_MAX_VALUE
> Then the order of timers is as follows (correct):
> - timerB
> - timerA
> But, if timerB sets another timer (say for timerB.timestamp + 1), then the
> order of timers will be:
> - timerB (timerB.timestamp)
> - timerA (BoundedWindow.TIMESTAMP_MAX_VALUE)
> - timerB (timerB.timestamp + 1)
> Which is not ordered by timestamp. The reason for this is that when the input
> watermark update is evaluated, the WatermarkManager,extractFiredTimers() will
> produce both timerA and timerB. That would be correct, but when timerB sets
> another timer, that breaks this.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
