[
https://issues.apache.org/jira/browse/FLINK-7406?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16196774#comment-16196774
]
ASF GitHub Bot commented on FLINK-7406:
---------------------------------------
Github user NicoK commented on a diff in the pull request:
https://github.com/apache/flink/pull/4509#discussion_r143233556
--- Diff:
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/consumer/RemoteInputChannel.java
---
@@ -331,37 +353,41 @@ public int getNumberOfAvailableBuffers() {
*/
@Override
public boolean notifyBufferAvailable(Buffer buffer) {
- checkState(isWaitingForFloatingBuffers.get(), "This channel
should be waiting for floating buffers.");
+ checkState(isWaitingForFloatingBuffers, "This channel should be
waiting for floating buffers.");
+ boolean needMoreBuffers;
synchronized (availableBuffers) {
// Important: the isReleased check should be inside the
synchronized block.
- if (isReleased.get() || availableBuffers.size() >=
senderBacklog.get()) {
- isWaitingForFloatingBuffers.set(false);
+ if (isReleased.get() || availableBuffers.size() >=
senderBacklog.get() + initialCredit) {
--- End diff --
1) The `senderBacklog` is only ever accessed under the
`synchronized(availableBuffers)`. I can't really think of a reason to drag it
out of there and keep it an `AtomicInteger` at the moment - what do you think
about making it a simple `int` instead?
2) Since we release extraneous floating buffers in `recycle()`, I guess,
`availableBuffers.size() > senderBacklog.get() + initialCredit` is not possible
here, is it? Should we `checkState()` this and otherwise ignore this case in
here?
> Implement Netty receiver incoming pipeline for credit-based
> -----------------------------------------------------------
>
> Key: FLINK-7406
> URL: https://issues.apache.org/jira/browse/FLINK-7406
> Project: Flink
> Issue Type: Sub-task
> Components: Network
> Reporter: zhijiang
> Assignee: zhijiang
> Fix For: 1.4.0
>
>
> This is a part of work for credit-based network flow control.
> Currently {{PartitionRequestClientHandler}} receives and reads
> {{BufferResponse}} from producer. It will request buffer from {{BufferPool}}
> for holding the message. If not got, the message is staged temporarily and
> {{autoread}} for channel is set false.
> For credit-based mode, {{PartitionRequestClientHandler}} can always get
> buffer from {{RemoteInputChannel}} for reading messages from producer.
> The related works are:
> * Add the backlog of producer in {{BufferResponse}} message structure
> * {{PartitionRequestClientHandler}} requests buffer from
> {{RemoteInputChannel}} directly
> * {{PartitionRequestClientHandler}} updates backlog for
> {{RemoteInputChannel}}, and it may trigger requests of floating buffers from
> {{BufferPool}}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
