[SPARK-1777] Pass "cached" blocks directly to disk if memory is not large enough

• Type: Improvement
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Fix Version/s: 1.1.0
• Component/s: Spark Core
• Assignee: Andrew Or
• Reporter: Patrick Wendell
• Created: 09/May/14 06:37
• Updated: 05/Nov/14 10:45
• Resolved: 27/Jul/14 23:08

Description

Currently in Spark we entirely unroll a partition and then check whether it will cause us to exceed the storage limit. This has an obvious problem - if the partition itself is enough to push us over the storage limit (and eventually over the JVM heap), it will cause an OOM.

This can happen in cases where a single partition is very large or when someone is running examples locally with a small heap.

https://github.com/apache/spark/blob/f6ff2a61d00d12481bfb211ae13d6992daacdcc2/core/src/main/scala/org/apache/spark/CacheManager.scala#L148

We should think a bit about the most elegant way to fix this - it shares some similarities with the external aggregation code.

A simple idea is to periodically check the size of the buffer as we are unrolling and see if we are over the memory limit. If we are we could prepend the existing buffer to the iterator and write that entire thing out to disk.

Issue Links

• is duplicated by: SPARK-1392 Local spark-shell Runs Out of Memory With Default Settings
• relates to: SPARK-1201 Do not materialize partitions whenever possible in BlockManager
• links to: [Github] Pull Request #1165 (andrewor14)
• links to: [Github] Pull Request #1892 (liyezhang556520)