flink-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Piotr Nowojski (Jira)" <j...@apache.org>
Subject [jira] [Updated] (FLINK-12912) Incorrect handling of task.checkpoint.alignment.max-size when one checkpoint subsumes another one
Date Thu, 29 Apr 2021 10:23:00 GMT

     [ https://issues.apache.org/jira/browse/FLINK-12912?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Piotr Nowojski updated FLINK-12912:
-----------------------------------
    Priority: Minor  (was: Major)

> Incorrect handling of task.checkpoint.alignment.max-size when one checkpoint subsumes
another one
> -------------------------------------------------------------------------------------------------
>
>                 Key: FLINK-12912
>                 URL: https://issues.apache.org/jira/browse/FLINK-12912
>             Project: Flink
>          Issue Type: Bug
>          Components: Runtime / Network
>    Affects Versions: 1.6.4, 1.7.2, 1.8.0
>            Reporter: Piotr Nowojski
>            Priority: Minor
>              Labels: stale-major
>
> {{BarrierBuffer#numQueuedBytes}} which is used to evaluate {{ task.checkpoint.alignment.max-size}}
limit, is not correctly handled if one checkpoint subsumes another one.
> The max size limit is checked against a sum of {{numQueuedBytes}} and {{bufferBlocker.getBytesBlocked()}}.
The {{getBytesBlocked}} keeps tracks of the alignment size of the only most latest checkpoint.
The bug is {{BarrierBuffer#releaseBlocksAndResetBarriers()}} method, where while handling
first subsumed checkpoint in the branch:
> {code:java}
> 		if (currentBuffered == null) {
> 			// common case: no more buffered data
> 			currentBuffered = bufferBlocker.rollOverReusingResources();
> 			if (currentBuffered != null) {
> 				currentBuffered.open();
> 			}
> 		}
> {code}
> we clear the {{bufferBlocker.getBytesBlocked()}} counter, while we do not update {{numQueuedBytes}}
counter. 
> For example when first checkpoint approached to 99.9% of max alignment size when it was
subsumed, due to this bug calculated alignment size drops to 0 bytes. For subsequent subsumed
checkpoints {{numQueuedBytes}} is correctly updated.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

Mime
View raw message