[
https://issues.apache.org/jira/browse/FLINK-4354?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15643637#comment-15643637
]
Zhijiang Wang edited comment on FLINK-4354 at 11/7/16 9:44 AM:
---------------------------------------------------------------
Hi [~till.rohrmann], for heartbeat interaction between {{TM}} and {{RM}}, I
have some issues need to be confirmed with you before implementation.
1. The current {{HeartbeatManagerImpl}} implements both {{HeartbeatManager}}
and {{HeartbeatTarget}} interfaces in order to test easily. I think we need
another HeartbeatManagerImpl that just implements {{HeartbeatManager}}
interface so can be used directly in different components. And every component
can implement the separate {{HeartbeatTarget}}.
2. For {{TM}} component, the {{HeartbeatManagerImpl}} can be constructed in
{{TaskManagerRunner}} (maybe not put in {{TaskManagerServices}}) and passed
into {{TaskExecutor}}.
3. The {{TM}} will create the {{HeartbeatListener}} and start the
{{HeartbeatManagerImpl}}.
4. When {{RM}} leader changes, the {{TM}} registers the new {{RM}}. If the
registration successes, {{TaskExecutorRegistrationSuccess}} should contain
{{ResourceID}} of {{RM}}, so the {{TM}} can create the {{HeartbeatTarget}} and
monitor it based on {{ResourceID}} and Gateway of RM.
5. For {{RM}}, when receive registration from {{TM}}, it will create the
{{HeartbeatTarget}} and monitor it based on {{ResourceID}} and Gateway of
{{TM}}.
{{RM}} will schedule a heartbeat request to all the monitored TMs.
6. {{TaskExecutorGateway}} should define the requestHeartbeat RPC method, and
{{ResourceManagerGateway}} should define the sendHeartbeat RPC method.
Do you think the above processes are feasible? I wish your professional
advices and them begin to implement this week.
was (Author: zjwang):
Hi [~till.rohrmann], for heartbeat interaction between {{TM}} and {{RM}}, I
have some issues need to be confirmed with you before implementation.
1. The current {{HeartbeatManagerImpl}} implements both {{HeartbeatManager}}
and {{HeartbeatTarget}} interfaces in order to test easily. I think we need
another HeartbeatManagerImpl that just implements HeartbeatManager interface so
can be used directly in different components. And every component can implement
the separate HeartbeatTarget.
2. For TM component, the HeartbeatManagerImpl can be constructed in
TaskManagerRunner (maybe not put in TaskManagerServices) and passed into
TaskExecutor.
3. The TM will create the HeartbeatListener and start the HeartbeatManagerImpl.
4. When RM leader changes, the TM registers the new RM. If the registration
successes, TaskExecutorRegistrationSuccess should contain ResourceID of RM, so
the TM can create the HeartbeatTarget and monitor it based on ResourceID and
Gateway of RM.
5. For RM, when receive registration from TM, it will create the
HeartbeatTarget and monitor it based on ResourceID and Gateway of TM.
RM will schedule a heartbeat request to all the monitored TMs.
6. TaskExecutorGateway should define the requestHeartbeat RPC method, and
ResourceManagerGateway should define the sendHeartbeat RPC method.
Do you think the above processes are feasible? I wish your professional
advices and them begin to implement this week.
> Implement TaskManager side of heartbeat from ResourceManager
> -------------------------------------------------------------
>
> Key: FLINK-4354
> URL: https://issues.apache.org/jira/browse/FLINK-4354
> Project: Flink
> Issue Type: Sub-task
> Components: Cluster Management
> Reporter: Zhijiang Wang
> Assignee: Zhijiang Wang
>
> The {{ResourceManager}} initiates heartbeat messages via the {{RmLeaderID}}.
> The {{TaskManager}} transmits its slot availability with each heartbeat. That
> way, the RM will always know about available slots.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
