[
https://issues.apache.org/jira/browse/APEXCORE-222?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15338980#comment-15338980
]
ASF GitHub Bot commented on APEXCORE-222:
-----------------------------------------
Github user vrozov commented on a diff in the pull request:
https://github.com/apache/apex-core/pull/350#discussion_r67635187
--- Diff:
engine/src/main/java/com/datatorrent/stram/engine/StreamingContainer.java ---
@@ -769,7 +769,11 @@ public void
processHeartbeatResponse(ContainerHeartbeatResponse rsp)
}
if (rsp.committedWindowId != lastCommittedWindowId) {
+
lastCommittedWindowId = rsp.committedWindowId;
+
+ bufferServer.purge(lastCommittedWindowId);
--- End diff --
While currently buffer server by itself does not make any assumption
regarding locality and communicates with its clients (publisher, subscriber,
controller) over network request API, Apex supports deployment of the buffer
server into the same container as the upstream operator/publisher only. As the
buffer server is not fault tolerant, and there is no plan to make it fault
tolerant, in reality, there is no flexibility to deploy buffer server into
another container other than the upstream operator container. Additionally, as
long as the buffer server continues to support purge request submitted over a
network, the flexibility to deploy buffer server in any container still can be
exercised once (or if) fault tolerance is supported by the buffer server.
> Delegate Buffer Server purge to StreamingContainer
> --------------------------------------------------
>
> Key: APEXCORE-222
> URL: https://issues.apache.org/jira/browse/APEXCORE-222
> Project: Apache Apex Core
> Issue Type: Improvement
> Reporter: Thomas Weise
> Assignee: Sandesh
>
> Currently the purge requests are sent to the buffer servers from the app
> master. This interaction exists parallel to the heartbeat protocol. Instead,
> the committed window ID that is propagated through the heartbeat response can
> be used in StreamingContainer to initiate the purge with the local buffer
> server, similar to how the committed callback on the operator occurs.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
