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] [Closed] (FLINK-10255) Standby Dispatcher locks submitted JobGraphs
Date Fri, 14 Sep 2018 13:18:00 GMT

     [ https://issues.apache.org/jira/browse/FLINK-10255?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Till Rohrmann closed FLINK-10255.
---------------------------------
    Resolution: Fixed

Fixed via 
1.7.0: 3e5d07ca349a7b010bc47d1cce9b9ad3208f55a6
1.6.1: b9c89d9a7af45f1b605e46c7d736c3bdc9b0d16f
1.5.4: 5a97f12c339ed3c0b6798c9fc0fd17910689099d

> Standby Dispatcher locks submitted JobGraphs
> --------------------------------------------
>
>                 Key: FLINK-10255
>                 URL: https://issues.apache.org/jira/browse/FLINK-10255
>             Project: Flink
>          Issue Type: Bug
>          Components: Distributed Coordination
>    Affects Versions: 1.5.3, 1.6.0, 1.7.0
>            Reporter: Till Rohrmann
>            Assignee: Till Rohrmann
>            Priority: Blocker
>              Labels: pull-request-available
>             Fix For: 1.6.1, 1.7.0, 1.5.4
>
>
> Currently, standby {{Dispatchers}} lock submitted {{JobGraphs}} which are added to the
{{SubmittedJobGraphStore}} if HA mode is enabled. Locking the {{JobGraphs}} can prevent their
cleanup leaving the system in an inconsistent state.
> The problem is that we recover in the {{SubmittedJobGraphListener#onAddedJobGraph}} callback
which is also called if don't have the leadership the newly added {{JobGraph}}. Recovering
the {{JobGraph}} currently locks the {{JobGraph}}. In case that the {{Dispatcher}} is not
the leader, then we won't start that job after its recovery. However, we also don't release
the {{JobGraph}} leaving it locked.
> There are two possible solutions to the problem. Either we check whether we are the leader
before recovering jobs or we say that recovering jobs does not lock them. Only if we can submit
the recovered job we lock them. The latter approach has the advantage that it follows a quite
similar code path as an initial job submission. Moreover, jobs are currently also recovered
at other places. In all these places we currently would need to release the {{JobGraphs}}
if we cannot submit the recovered {{JobGraph}} (e.g. {{Dispatcher#grantLeadership}}).
> An extension of the first solution could be to stop the {{SubmittedJobGraphStore}} while
the {{Dispatcher}} is not the leader. Then we would have to make sure that no concurrent callback
from the {{SubmittedJobGraphStore#SubmittedJobGraphListener}} can be executed after revoking
leadership from the {{Dispatcher}}.



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

Mime
View raw message