squito commented on a change in pull request #23688: [SPARK-25035][Core]
Avoiding memory mapping at disk-stored blocks replication
URL: https://github.com/apache/spark/pull/23688#discussion_r255698221
##########
File path: core/src/main/scala/org/apache/spark/storage/BlockManager.scala
##########
@@ -435,24 +606,8 @@ private[spark] class BlockManager(
// Note this is all happening inside the netty thread as soon as it
reads the end of the
// stream.
channel.close()
- // TODO SPARK-25035 Even if we're only going to write the data to disk
after this, we end up
- // using a lot of memory here. We'll read the whole file into a regular
- // byte buffer and OOM. We could at least read the tmp file as a
stream.
- val buffer = securityManager.getIOEncryptionKey() match {
- case Some(key) =>
- // we need to pass in the size of the unencrypted block
- val blockSize = channel.getCount
- val allocator = level.memoryMode match {
- case MemoryMode.ON_HEAP => ByteBuffer.allocate _
- case MemoryMode.OFF_HEAP => Platform.allocateDirectBuffer _
- }
- new EncryptedBlockData(tmpFile, blockSize, conf,
key).toChunkedByteBuffer(allocator)
-
- case None =>
- ChunkedByteBuffer.fromFile(tmpFile)
- }
- putBytes(blockId, buffer, level)(classTag)
- tmpFile.delete()
+ val blockSize = channel.getCount
Review comment:
updated the comment at the beginning of `onComplete` here
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
