[
https://issues.apache.org/jira/browse/BEAM-8577?focusedWorklogId=365864&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-365864
]
ASF GitHub Bot logged work on BEAM-8577:
----------------------------------------
Author: ASF GitHub Bot
Created on: 03/Jan/20 14:40
Start Date: 03/Jan/20 14:40
Worklog Time Spent: 10m
Work Description: cchepelov commented on issue #10027: [BEAM-8577]
Initialize FileSystems during Coder deserialization in Re…
URL: https://github.com/apache/beam/pull/10027#issuecomment-570590340
I've been hit by the same problem in a
- scio 0.8.0-beta2 (cc @nevillelyh @jto )
- scala 2.12 (.10)
- beam 2.16.0
- flink 1.8.3 (or 1.9.1)
job while trying to use the flink runner on a cluster. Same issue as
described in BEAM-8577, the root cause was the reshuffle and uninitialized
(BEAM) FileSystems registry across the ResourceIdCoder.
Backporting @dmvk's patch to beam 2.16.0 (including persuading gradle into
making a local Ivy publication to get picked up by SBT) successfully led to
reliably hit another, unrelated problem. Would love to see this PR go through!
Thank you
----------------------------------------------------------------
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: 365864)
Time Spent: 50m (was: 40m)
> FileSystems may have not be initialized during ResourceId deserialization
> -------------------------------------------------------------------------
>
> Key: BEAM-8577
> URL: https://issues.apache.org/jira/browse/BEAM-8577
> Project: Beam
> Issue Type: Bug
> Components: runner-flink
> Affects Versions: 2.16.0
> Reporter: David Morávek
> Assignee: David Morávek
> Priority: Major
> Time Spent: 50m
> Remaining Estimate: 0h
>
> - FileSystems use static registration using
> *FileSystems#setDefaultPipelineOptions* method.
> - *#setDefaultPipelineOptions* is called either when deserializaing
> SerializablePipelineOptions or during opening of various beam operators.
> - *FileIO#matchAll* is expanded using *Reshuffle.viaRandomKey()*.
> - Reshuffle is implemented using *.rebalance*, that doesn't have a
> "RichFunction" lifecycle, so we need to find another way to register
> FileSystems, as the deserialization may happen before other "rich operators"
> get executed on particular task manager.
> This results in random pipeline fails as the task assignment is not
> deterministic.
> We can workaround this, by registering FileSystems during coder
> deserialization.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
