flink-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Till Rohrmann (JIRA)" <j...@apache.org>
Subject [jira] [Resolved] (FLINK-8740) Job-level metrics lost during job re-submission in HA mode
Date Wed, 28 Mar 2018 15:28:00 GMT

     [ https://issues.apache.org/jira/browse/FLINK-8740?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Till Rohrmann resolved FLINK-8740.
----------------------------------
    Resolution: Fixed

Fixed via
master: 9a31deae7ebdbda5a78f19ae8967e6b9d35834a4
1.5.0: 237563215fa9d8f51fc9387c217e0fafed095fca

> Job-level metrics lost during job re-submission in HA mode
> ----------------------------------------------------------
>
>                 Key: FLINK-8740
>                 URL: https://issues.apache.org/jira/browse/FLINK-8740
>             Project: Flink
>          Issue Type: Bug
>          Components: JobManager
>    Affects Versions: 1.4.0
>            Reporter: Joshua DeWald
>            Assignee: Till Rohrmann
>            Priority: Blocker
>             Fix For: 1.5.0
>
>
> When Flink is running in High Availability and a leader re-election occurs to the same
job manager, the job is unable to register the job-level metrics due to a name collision. 
> This may occur even if a different Job Manager is elected, but as it is a local JobManagerMetricsGroup
which spits out the error, that is unlikely the case.
>  
> *Expected Behavior*
> When a job is forced to re-submit due to Job Manager re-election, job-level metrics
should be available in the new instance of the job (uptime, checkpoints size, checkpoint duration,
etc)
> *Actual Behavior*
> When job gets re-submitted, it is unable to register job-level metrics due to collision
in the JobManagerMetricGroup, which leads to situation where even though job is running the
metrics around checkpoints and uptime are not available
> *Steps to reproduce*
>  # Start up Flink in HA mode using ZooKeeper, single node is fine
>  # Submit a job to the cluster
>  # Stop and restart ZooKeeper
>  # In Job Manager logs you will see the following errors:
>  # 
> {noformat}
> 79043 2018-02-19 21:58:15,928 WARN org.apache.flink.metrics.MetricGroup - Name collision:
Group already contains a Metric with the name 'totalNumberOfCheckpoints'. Metric will not
be reported....
> 79044 2018-02-19 21:58:15,928 WARN org.apache.flink.metrics.MetricGroup - Name collision:
Group already contains a Metric with the name 'numberOfInProgressCheckpoints'. Metric will
not be reported....
> 79045 2018-02-19 21:58:15,928 WARN org.apache.flink.metrics.MetricGroup - Name collision:
Group already contains a Metric with the name 'numberOfCompletedCheckpoints'. Metric will
not be reported....{noformat}
> *Proposed Solution*
> I suspect that there may be other related issues than just the metrics, but a code change
that seems to fix the issue is that, during recovery, to remove the existing registered Job
Metrics:
> {code:java}
> if (isRecovery) {
>    log.info(s"Removing metrics for $jobId, new will be added during recover")
>    jobManagerMetricGroup.removeJob(jobId)
> }{code}
> I'd be happy to submit this in a PR if that is acceptable to open up the discussion,
but I am not sure the consequences of not closing the previous JMMG or perhaps simply not
re-registering job-level metrics during recovery. Doing this would seem to entail informing
lower levels about the recovery.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Mime
View raw message