[
https://issues.apache.org/jira/browse/TEZ-4075?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16944252#comment-16944252
]
Rajesh Balamohan commented on TEZ-4075:
---------------------------------------
Thanks [~sseth] for the review comments.
Closing the compressed stream and removing the checksum is a great idea.
However, in corner cases the left over data is exceeding the buffer size
causing overflow issues.
Instead of using compressed stream, tried using uncompressed stream (as the
data checks are done in uncompressed manner anyways). But this also leads to a
corner case issue, as FSD*Stream
is backed by buffered stream of 4096 array. It hits the limit during buffer
close (i.e at the time of out.close, it tries to flush its contents and causes
overflow). Along with this, we need to fix file close markers and comp/decom
bytes counters during corner case scearnios.
Given the corner case conditions, it would be good to split the patch into 2.
1) re-enable payload via events in this patch. 2) Aim towards reducing the file
copy call in subsequent jira.
I will create a PR as mentioned in option 1.
> Tez: Reimplement tez.runtime.transfer.data-via-events.enabled
> -------------------------------------------------------------
>
> Key: TEZ-4075
> URL: https://issues.apache.org/jira/browse/TEZ-4075
> Project: Apache Tez
> Issue Type: Bug
> Reporter: Gopal Vijayaraghavan
> Assignee: Richard Zhang
> Priority: Major
> Attachments: TEZ-4075.10.patch, TEZ-4075.15.patch, TEZ-4075.16.patch,
> Tez-4075.5.patch, Tez-4075.8.patch
>
>
> This was factored out by TEZ-2196, which does skip buffers for 1-partition
> data exchanges (therefore goes to disk directly).
> {code}
> if (shufflePayload.hasData()) {
> shuffleManager.addKnownInput(shufflePayload.getHost(),
> DataProto dataProto = shufflePayload.getData();
> shufflePayload.getPort(), srcAttemptIdentifier, srcIndex);
> FetchedInput fetchedInput =
> inputAllocator.allocate(dataProto.getRawLength(),
> dataProto.getCompressedLength(), srcAttemptIdentifier);
> moveDataToFetchedInput(dataProto, fetchedInput, hostIdentifier);
> shuffleManager.addCompletedInputWithData(srcAttemptIdentifier,
> fetchedInput);
> } else {
> shuffleManager.addKnownInput(shufflePayload.getHost(),
> shufflePayload.getPort(), srcAttemptIdentifier, srcIndex);
> }
> {code}
> got removed in
> https://github.com/apache/tez/commit/1ba1f927c16a1d7c273b6cd1a8553e5269d1541a
> It would be better to buffer up the 512Byte limit for the event size before
> writing to disk, since creating a new file always incurs disk traffic, even
> if the file is eventually being served out of the buffer cache.
> The total overhead of receiving an event, then firing an HTTP call to fetch
> the data etc adds approx 100-150ms to a query - the data xfer through the
> event will skip the disk entirely for this & also remove the extra IOPS
> incurred.
> This channel is not suitable for large-scale event transport, but
> specifically the workload here deals with 1-row control tables which consume
> more bandwidth with HTTP headers and hostnames than the 93 byte payload.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
