[
https://issues.apache.org/jira/browse/FLINK-4455?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15453760#comment-15453760
]
ASF GitHub Bot commented on FLINK-4455:
---------------------------------------
GitHub user tillrohrmann opened a pull request:
https://github.com/apache/flink/pull/2449
[FLINK-4455] [FLINK-4424] [networkenv] Make NetworkEnvironment independent
of ActorGateway and JobManager association
Makes the NetworkEnvironment independent of the JobManager association.
This means that the
NetworkEnvironment and with it the ConnectionManager is started before the
TaskManager actor
is executed. Furthermore, the ConnectionManager keeps running even in case
of a JobManager
disassocation. In the wake of the remodelling this behaviour, the
PartitionStateChecker and
the ResultPartitionConsumableNotifier which depend on the JobManager
association were moved
out of the NetworkEnvironment. They are now contained in the
SlotEnvironment which will be
set up when the TaskManager connects to a JobManager. The SlotEnvironment
contains all
information related to the associated JobManager. Since all slots are
implicitly associated
with the JobManager which is the leader, we only create one SlotEnvironment
which is shared
by all Tasks.
Introduce SlotEnvironment to accommodate the PartitionStateChecker and
ResultPartitionConsumableNotifier
Remove the PartitionStateChecker and the ResultPartitionConsumableNotifier
from the
NetworkEnvironment. Start the NetworkEnvironment when the TaskManager
components are
created. Keep the NetworkEnvironment running also when the JobManager is
disassociated.
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/tillrohrmann/flink FLINK-4455
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/flink/pull/2449.patch
To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:
This closes #2449
----
commit 96b1772d4581365eeec6614827b36f76a83ba5d0
Author: Till Rohrmann <trohrm...@apache.org>
Date: 2016-08-31T07:33:46Z
[FLINK-4455] [FLINK-4424] [networkenv] Make NetworkEnvironment independent
of ActorGateway and JobManager association
Makes the NetworkEnvironment independent of the JobManager association.
This means that the
NetworkEnvironment and with it the ConnectionManager is started before the
TaskManager actor
is executed. Furthermore, the ConnectionManager keeps running even in case
of a JobManager
disassocation. In the wake of the remodelling this behaviour, the
PartitionStateChecker and
the ResultPartitionConsumableNotifier which depend on the JobManager
association were moved
out of the NetworkEnvironment. They are now contained in the
SlotEnvironment which will be
set up when the TaskManager connects to a JobManager. The SlotEnvironment
contains all
information related to the associated JobManager. Since all slots are
implicitly associated
with the JobManager which is the leader, we only create one SlotEnvironment
which is shared
by all Tasks.
Introduce SlotEnvironment to accommodate the PartitionStateChecker and
ResultPartitionConsumableNotifier
Remove the PartitionStateChecker and the ResultPartitionConsumableNotifier
from the
NetworkEnvironment. Start the NetworkEnvironment when the TaskManager
components are
created. Keep the NetworkEnvironment running also when the JobManager is
disassociated.
----
> Replace ActorGateways in NetworkEnvironment by interfaces
> ---------------------------------------------------------
>
> Key: FLINK-4455
> URL: https://issues.apache.org/jira/browse/FLINK-4455
> Project: Flink
> Issue Type: Improvement
> Components: Network, TaskManager
> Reporter: Till Rohrmann
> Assignee: Till Rohrmann
>
> The {{NetworkEnvironment}} communicates with the outside world
> ({{TaskManager}} and {{JobManager}}) via {{ActorGateways}}. This bakes in the
> dependency on actors.
> In terms of modularization and an improved abstraction (especially wrt
> Flip-6) I propose to replace the {{ActorGateways}} by interfaces which
> exposes the required methods. The current implementation would then simply
> wrap the method calls in messages and send them via the {{ActorGateway}} to
> the recipient.
> In Flip-6 the {{JobMaster}} and the {{TaskExecutor}} could simply implement
> these interfaces as part of their RPC contract.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
