Github user andrewor14 commented on a diff in the pull request: https://github.com/apache/spark/pull/1165#discussion_r14723210 --- Diff: core/src/main/scala/org/apache/spark/CacheManager.scala --- @@ -142,10 +151,76 @@ private[spark] class CacheManager(blockManager: BlockManager) extends Logging { * to the BlockManager as an iterator and expect to read it back later. This is because * we may end up dropping a partition from memory store before getting it back, e.g. * when the entirety of the RDD does not fit in memory. */ - val elements = new ArrayBuffer[Any] - elements ++= values - updatedBlocks ++= blockManager.put(key, elements, storageLevel, tellMaster = true) - elements.iterator.asInstanceOf[Iterator[T]] + --- End diff -- The logic here actually hasn't changed (the code was just moved elsewhere). For the `MEMORY_ONLY` case, if we can't be sure that there is enough memory to unroll the block, we simply drop it on the spot. I don't believe this is a violation of the StorageLevel contract insofar as old blocks cached as `MEMORY_ONLY` can also be dropped from the cache. The only difference here is that we drop it earlier (before it has the potential to cause an OOM). Among the options you listed, the last thing we want is (3), which unfortunately is what happens in the existing code. It's true that if we our estimate is too high, we may drop a block that can otherwise be cached. However, the whole point of this is to guarantee memory safety (i.e. to avoid OOMs), and to achieve that we need to err on the safe side.
--- 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 infrastruct...@apache.org or file a JIRA ticket with INFRA. ---