[
https://issues.apache.org/jira/browse/FLINK-5645?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15878262#comment-15878262
]
ASF GitHub Bot commented on FLINK-5645:
---------------------------------------
Github user tillrohrmann commented on a diff in the pull request:
https://github.com/apache/flink/pull/3377#discussion_r102464984
--- Diff:
flink-runtime/src/main/java/org/apache/flink/runtime/executiongraph/ExecutionGraph.java
---
@@ -1292,10 +1292,28 @@ public boolean updateState(TaskExecutionState
state) {
}
return true;
case CANCELED:
- attempt.cancelingComplete();
+ AccumulatorSnapshot acc1 =
state.getAccumulators();
+ Map<String, Accumulator<?, ?>> userAcc1
= null;
+ if (acc1 != null) {
+ try {
+ userAcc1 =
acc1.deserializeUserAccumulators(userClassLoader);
+ } catch (Exception e) {
+ LOG.error("Failed to
deserialize final accumulator results.", e);
+ }
+ }
+ attempt.cancelingComplete(userAcc1,
state.getIOMetrics());
return true;
case FAILED:
-
attempt.markFailed(state.getError(userClassLoader));
+ AccumulatorSnapshot acc2 =
state.getAccumulators();
+ Map<String, Accumulator<?, ?>> userAcc2
= null;
+ if (acc2 != null) {
+ try {
+ userAcc2 =
acc2.deserializeUserAccumulators(userClassLoader);
+ } catch (Exception e) {
+ LOG.error("Failed to
deserialize final accumulator results.", e);
+ }
+ }
--- End diff --
Can this deserialization logic be factored out into a method?
> IOMetrics transfer through ExecGraph does not work for failed jobs
> ------------------------------------------------------------------
>
> Key: FLINK-5645
> URL: https://issues.apache.org/jira/browse/FLINK-5645
> Project: Flink
> Issue Type: Bug
> Components: Metrics, Webfrontend
> Affects Versions: 1.2.0, 1.3.0
> Reporter: Chesnay Schepler
> Assignee: Chesnay Schepler
> Priority: Blocker
>
> While a job is running the web-interface polls IO metrics using the
> MetricFetcher/MetricQueryService. For finished jobs these metrics are stored
> in the ExecutionGraph; each vertex has an IOMetrics object; basically a
> snapshot taken when the task was finished.
> For completed jobs this mechanism is working. However, for failed jobs it
> doesn't; in this case last fetched metrics are displayed.
> This isn't intended behavior and should be fixed.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
