Github user JoshRosen commented on a diff in the pull request:
https://github.com/apache/spark/pull/2844#discussion_r19063336
--- Diff:
core/src/main/scala/org/apache/spark/broadcast/TorrentBroadcast.scala ---
@@ -76,23 +87,20 @@ private[spark] class TorrentBroadcast[T: ClassTag](
* @return number of blocks this broadcast variable is divided into
*/
private def writeBlocks(): Int = {
- // For local mode, just put the object in the BlockManager so we can
find it later.
- SparkEnv.get.blockManager.putSingle(
- broadcastId, _value, StorageLevel.MEMORY_AND_DISK, tellMaster =
false)
-
- if (!isLocal) {
- val blocks = TorrentBroadcast.blockifyObject(_value)
- blocks.zipWithIndex.foreach { case (block, i) =>
- SparkEnv.get.blockManager.putBytes(
- BroadcastBlockId(id, "piece" + i),
- block,
- StorageLevel.MEMORY_AND_DISK_SER,
- tellMaster = true)
- }
- blocks.length
- } else {
- 0
+ // Store a copy of the broadcast variable in the driver so that tasks
run on the driver
+ // do not create a duplicate copy of the broadcast variable's value.
+ SparkEnv.get.blockManager.putSingle(broadcastId, _value,
StorageLevel.MEMORY_AND_DISK,
+ tellMaster = false)
--- End diff --
The reason for this store is to avoid creating two copies of `_value` in
the driver. If we serialize and deserialize a broadcast variable on the driver
and then attempt to access its value, then without this code we will end up
going through the regular de-chunking code path, which will cause us to
deserialize the serialized copy of `_value` and waste memory.
I believe that this serialization and deserialization can take place when
tasks are run in local mode, since we still serialize tasks in order to help
users be aware of serialization issues that would impact them if they moved to
a cluster. This complexity is another reason why I'm in favor of just
scrapping all local-mode special-casing and configuring Spark to use a dummy
LocalBroadcastFactory for local mode instead of whichever setting the user
specified. That would be a larger, more-invasive change, which is why I opted
for the simpler fix here.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
