otterc commented on a change in pull request #32287:
URL: https://github.com/apache/spark/pull/32287#discussion_r622721640
##########
File path:
core/src/main/scala/org/apache/spark/storage/ShuffleBlockFetcherIterator.scala
##########
@@ -261,14 +282,36 @@ final class ShuffleBlockFetcherIterator(
results.put(new SuccessFetchResult(BlockId(blockId),
infoMap(blockId)._2,
address, infoMap(blockId)._1, buf, remainingBlocks.isEmpty))
logDebug("remainingBlocks: " + remainingBlocks)
+ enqueueDeferredFetchRequestIfNecessary()
}
}
logTrace(s"Got remote block $blockId after
${Utils.getUsedTimeNs(startTimeNs)}")
}
override def onBlockFetchFailure(blockId: String, e: Throwable): Unit = {
logError(s"Failed to get block(s) from
${req.address.host}:${req.address.port}", e)
- results.put(new FailureFetchResult(BlockId(blockId),
infoMap(blockId)._2, address, e))
+ remainingBlocks -= blockId
+ val (size, mapIndex) = infoMap(blockId)
+ e match {
+ // Catching OOM and do something based on it is only a workaround
for handling the
+ // Netty OOM issue, which is not the best way towards memory
management. We can
+ // get rid of it when we find a way to manage Netty's memory
precisely.
+
+ // Ensure the Netty memory is at least enough for serving only one
block to avoid
+ // the endless retry. And since the Netty memory is shared among
shuffle, rpc, etc,
+ // modules, we use "1.5" for the overhead concern.
+ case _: OutOfDirectMemoryError if
PlatformDependent.maxDirectMemory() > ( 1.5 * size) =>
+ if (isNettyOOMOnShuffle.compareAndSet(false, true)) {
+ // The fetcher can fail remaining blocks in batch for the same
error. So we only
+ // log the warning once to avoid flooding the logs.
+ logWarning(s"Netty OOM happens, will retry the failed blocks")
+ }
+ deferredBlocks += blockId
Review comment:
Sorry @Ngone51, I made a mistake here. The updates to the
`deferredBlocks` will be done by a single netty event-loop not different netty
event-loop so moving it inside `ShuffleBlockFetcherIterator.this.synchronized`
should not be necessary.
When `shuffleClient.fetchBlocks(....)` is called, it creates a channel and
at that time a netty event loop gets associated with this channel. This is the
event loop which will always be the one invoking
`blockFetchingListener.onFailure()` or `blockFetchingListener.onSuccess()` for
all the blocks in the request.
In the `onBlockFetchSuccess()` it uses
`ShuffleBlockFetcherIterator.this.synchronized` but that is because I think it
is accessing the `var isZombie` which can also be modified by the task thread.
cc @attilapiros
--
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]
