[
https://issues.apache.org/jira/browse/BEAM-7995?focusedWorklogId=300736&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-300736
]
ASF GitHub Bot logged work on BEAM-7995:
----------------------------------------
Author: ASF GitHub Bot
Created on: 24/Aug/19 16:49
Start Date: 24/Aug/19 16:49
Worklog Time Spent: 10m
Work Description: lukecwik commented on issue #9364: BEAM-7995: python
PGBKCVOperation using wrong timestamp
URL: https://github.com/apache/beam/pull/9364#issuecomment-524565147
PGBKCV = PartialGroupByKeyCombineValues
represents the mappers side lifted combine
https://s.apache.org/beam-runner-api-combine-model
On Fri, Aug 23, 2019 at 4:23 PM Kyle Weaver <notificati...@github.com>
wrote:
> This is only tangential to this PR, but while we're here, can someone tell
> me what "PGBKCV" stands for? 😅 I'd like to see this documented.
> https://issues.apache.org/jira/browse/BEAM-8087
>
> —
> You are receiving this because you were mentioned.
> Reply to this email directly, view it on GitHub
>
<https://github.com/apache/beam/pull/9364?email_source=notifications&email_token=ACM4V3ABPBSYAVK2C5B74YTQGBWPDA5CNFSM4IMM4NOKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD5BRSRA#issuecomment-524491076>,
> or mute the thread
>
<https://github.com/notifications/unsubscribe-auth/ACM4V3EDA54H42UT7GN323LQGBWPDANCNFSM4IMM4NOA>
> .
>
----------------------------------------------------------------
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:
us...@infra.apache.org
Issue Time Tracking
-------------------
Worklog Id: (was: 300736)
Time Spent: 1h 40m (was: 1.5h)
> IllegalStateException: TimestampCombiner moved element from to earlier time
> in Python
> -------------------------------------------------------------------------------------
>
> Key: BEAM-7995
> URL: https://issues.apache.org/jira/browse/BEAM-7995
> Project: Beam
> Issue Type: Bug
> Components: sdk-py-core
> Reporter: Hai Lu
> Assignee: Hai Lu
> Priority: Major
> Time Spent: 1h 40m
> Remaining Estimate: 0h
>
> I'm looking into a bug I found internally when using Beam portable API
> (Python) on our own Samza runner.Â
> Â
> The pipeline looks something like this:
> Â
> Â Â (p
> Â Â Â | 'read' >> ReadFromKafka(cluster="tracking", topic="PageViewEvent")
> Â Â Â | 'transform' >> beam.Map(lambda event: process_event(event))
> Â Â Â | 'window' >> beam.WindowInto(FixedWindows(15))
> Â Â Â | 'group' >>Â *beam.CombinePerKey(beam.combiners.CountCombineFn())*
> Â Â Â ...
> Â
> The problem comes from the combiners which cause the following exception on
> Java side:
> Â
> Caused by: java.lang.IllegalStateException: TimestampCombiner moved element
> from 2019-08-15T03:34:*45.000*Z to earlier time 2019-08-15T03:34:*44.999*Z
> for window [2019-08-15T03:34:30.000Z..2019-08-15T03:34:*45.000*Z)
> Â Â at
> org.apache.beam.runners.core.WatermarkHold.shift(WatermarkHold.java:117)
> Â Â at
> org.apache.beam.runners.core.WatermarkHold.addElementHold(WatermarkHold.java:154)
> Â Â at
> org.apache.beam.runners.core.WatermarkHold.addHolds(WatermarkHold.java:98)
> Â Â at
> org.apache.beam.runners.core.ReduceFnRunner.processElement(ReduceFnRunner.java:605)
> Â Â at
> org.apache.beam.runners.core.ReduceFnRunner.processElements(ReduceFnRunner.java:349)
> Â Â at
> org.apache.beam.runners.core.GroupAlsoByWindowViaWindowSetNewDoFn.processElement(GroupAlsoByWindowViaWindowSetNewDoFn.java:136)
> Â
> The exception happens hereÂ
> [https://github.com/apache/beam/blob/master/runners/core-java/src/main/java/org/apache/beam/runners/core/WatermarkHold.java#L116]
> Â when we check the shifted timestamp to ensure it's before the timestamp.
> Â
> Â Â if (shifted.isBefore(timestamp)) {
> Â Â Â throw new IllegalStateException(
> Â Â Â Â Â String.format(
> Â Â Â Â Â Â Â "TimestampCombiner moved element from %s to earlier time %s for
> window %s",
> Â Â Â Â Â Â Â BoundedWindow.formatTimestamp(timestamp),
> Â Â Â Â Â Â Â BoundedWindow.formatTimestamp(shifted),
> Â Â Â Â Â Â Â window));
> Â Â }
> Â
> As you can see from the exception, the "shifted" is "XXX 44.999" while the
> "timestamp" is "XXX 45.000". The "44.999" is coming fromÂ
> [TimestampCombiner.END_OF_WINDOW|https://github.com/apache/beam/blob/master/sdks/java/core/src/main/java/org/apache/beam/sdk/transforms/windowing/TimestampCombiner.java#L116]:
> Â
> Â Â @Override
> Â Â public Instant merge(BoundedWindow intoWindow, Iterable<? extends
> Instant> mergingTimestamps) {
> Â Â Â return intoWindow.maxTimestamp();
> Â Â }
> Â
> where intoWindow.maxTimestamp() is:
> Â
> Â /** Returns the largest timestamp that can be included in this window. */
> Â @Override
> Â public Instant maxTimestamp() {
> Â Â Â *// end not inclusive*
>   return *end.minus(1)*;
> Â }
> Â
> Hence, the "44.*999*".Â
> Â
> And the "45.000" comes from the Python side when the combiner output results
> as pre GBK operation:Â
> [operations.py#PGBKCVOperation#output_key|https://github.com/apache/beam/blob/master/sdks/python/apache_beam/runners/worker/operations.py#L889]
> Â
> Â Â if windows is 0:
> Â Â Â self.output(_globally_windowed_value.with_value((key, value)))
> Â Â else:
>    self.output(WindowedValue((key, value), *windows[0].end*, windows))
> Â
> Here when we generate the window value, the timestamp is assigned to the
> closed interval end (45.000) as opposed to open interval end (44.999)
> Â
> Clearly the "end of window" definition is a bit inconsistent across Python
> and Java. I'm yet to try this on other runner so not sure whether this is
> only an issue for our Samza runner. I tend to think this is a bug but would
> like to confirm with you. If this has not been an issue for other runners,
> where did I potentially do wrong.
--
This message was sent by Atlassian Jira
(v8.3.2#803003)
