[
https://issues.apache.org/jira/browse/FLINK-4894?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15611338#comment-15611338
]
ASF GitHub Bot commented on FLINK-4894:
---------------------------------------
Github user uce commented on a diff in the pull request:
https://github.com/apache/flink/pull/2690#discussion_r85299971
--- Diff:
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/api/writer/RecordWriter.java
---
@@ -108,15 +108,25 @@ private void sendToTarget(T record, int
targetChannel) throws IOException, Inter
synchronized (serializer) {
SerializationResult result =
serializer.addRecord(record);
+
while (result.isFullBuffer()) {
Buffer buffer = serializer.getCurrentBuffer();
if (buffer != null) {
- writeBuffer(buffer, targetChannel,
serializer);
+ writeAndClearBuffer(buffer,
targetChannel, serializer);
+
+ // If this was a full record, we are
done. Not breaking
+ // out of the loop at this point will
lead to another
+ // buffer request before breaking out
(that would not be
+ // a problem per se, but it can lead to
stalls in the
+ // pipeline).
+ if (result.isFullRecord()) {
+ break;
--- End diff --
Yes, exactly.
> Don't block on buffer request after broadcastEvent
> ---------------------------------------------------
>
> Key: FLINK-4894
> URL: https://issues.apache.org/jira/browse/FLINK-4894
> Project: Flink
> Issue Type: Improvement
> Reporter: Ufuk Celebi
> Assignee: Ufuk Celebi
>
> After broadcasting an event (like the checkpoint barrier), the record writer
> might block on a buffer request although that buffer will only be needed on
> the next write on that channel.
> Instead of assuming that each serializer has a buffer set, we can change the
> logic in the writer to request the buffer when it requires one.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
