jiangxb1987 commented on a change in pull request #29855:
URL: https://github.com/apache/spark/pull/29855#discussion_r500528147
##########
File path:
common/network-common/src/main/java/org/apache/spark/network/server/TransportRequestHandler.java
##########
@@ -181,6 +182,17 @@ public void onFailure(Throwable e) {
private void processStreamUpload(final UploadStream req) {
assert (req.body() == null);
try {
+ // Retain the original metadata buffer, since it will be used during the
invocation of
+ // this method. Will be released later.
+ req.meta.retain();
+ // Make a copy of the original metadata buffer. In benchmark, we noticed
that
+ // we cannot respond the original metadata buffer back to the client,
otherwise
+ // in cases where multiple concurrent shuffles are present, a wrong
metadata might
+ // be sent back to client. This is related to the eager release of the
metadata buffer,
+ // i.e., we always release the original buffer by the time the
invocation of this
+ // method ends, instead of by the time we respond it to the client. This
is necessary,
+ // otherwise we start seeing memory issues very quickly in benchmarks.
+ ByteBuffer meta = cloneBuffer(req.meta.nioByteBuffer());
Review comment:
Are we really adding a lot of memory pressure by not reusing the
callback? I'm expecting a size of ~10KB for each BlockPushCallback, so it
should be fine if we have hundreds or thousands of them. If the size is well
over hundreds of KBs then I'd assume something went wrong.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
