[
https://issues.apache.org/jira/browse/TWILL-190?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15437601#comment-15437601
]
ASF GitHub Bot commented on TWILL-190:
--------------------------------------
Github user chtyim commented on a diff in the pull request:
https://github.com/apache/twill/pull/4#discussion_r76318343
--- Diff:
twill-yarn/src/main/java/org/apache/twill/internal/appmaster/ApplicationMasterService.java
---
@@ -268,20 +270,33 @@ public void acquired(List<? extends
ProcessLauncher<YarnContainerInfo>> launcher
@Override
public void completed(List<YarnContainerStatus> completed) {
for (YarnContainerStatus status : completed) {
+ handleCompleted(completed);
ids.remove(status.getContainerId());
}
}
};
- runningContainers.stopAll();
-
- // Poll for 5 seconds to wait for containers to stop.
- int count = 0;
- while (!ids.isEmpty() && count++ < 5) {
- amClient.allocate(0.0f, handler);
- TimeUnit.SECONDS.sleep(1);
- }
+ // Handle heartbeats during shutdown because
runningContainers.stopAll() waits until
+ // handleCompleted() is called for every stopped runnable
+ ExecutorService stopPoller =
Executors.newSingleThreadExecutor(Threads.createDaemonThreadFactory("stopPoller"));
+ stopPoller.execute(new Runnable() {
+ @Override
+ public void run() {
+ while (!ids.isEmpty()) {
+ try {
+ amClient.allocate(0.0f, handler);
+ TimeUnit.SECONDS.sleep(1);
--- End diff --
Should check if `ids` is already emptied before sleeping, since the call
the `allocate` may already have the ids emptied by the handler and we don't
have the sleep for an extra second for that.
> Restart of a TwillRunnable does not wait for the runnable to stop
> -----------------------------------------------------------------
>
> Key: TWILL-190
> URL: https://issues.apache.org/jira/browse/TWILL-190
> Project: Apache Twill
> Issue Type: Bug
> Components: core, yarn
> Affects Versions: 0.6.0-incubating, 0.7.0-incubating
> Reporter: Poorna Chandra
> Assignee: Poorna Chandra
> Fix For: 0.8.0
>
>
> Today when a TwillRunnable is restarted, the call sends a stop message to the
> TwillRunnable, and then starts new TwillRunnable without waiting for the
> stopping runnable to finish stopping.
> This can leave a non-responding TwillRunnable container running, and can lead
> to issues like two TwillRunnables with same instance id running at the same
> time.
> We should kill the containers that don't respond to stop message.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
