[DRILL-4823] Fix OOM while trying to prune partitions with reasonable data size

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.6.0
• Fix Version/s: 1.9.0
• Component/s: Functions - Drill, Query Planning & Optimization
• Assignee: Arina Ielchiieva
• Reporter: Arina Ielchiieva
• Reviewer: Rahul Kumar Challapalli
• Created: 03/Aug/16 09:30
• Updated: 20/Mar/17 21:56
• Resolved: 05/Nov/16 01:08

Description

Example query:

select  '/'||dir0||'/'||dir1||'/'||dir2||'/'||dir3 , count(*)
  FROM dfs.`/path/to/parquet/files`
  WHERE ('/'||dir0||'/'||dir1||'/'||dir2||'/'||dir3)  IN ( 
  '/2015/11/30', 
  '//2015/09/01',
  '/2015/09/02', 
  '/2015/09/03',
  '/2015/09/04',
  '/2015/09/09',  
  '/2016/03/30'
  )
  group by   '/'||dir0||'/'||dir1||'/'||dir2||'/'||dir3
  order by 1;

Error:
OOM while trying to prune partitions:

org.apache.drill.exec.exception.OutOfMemoryException: Unable to allocate buffer of size 256 due to memory limit. Current allocation: 5242880
	at org.apache.drill.exec.memory.BaseAllocator.buffer(BaseAllocator.java:216) ~[drill-memory-base-1.6.0.jar:1.6.0]
	at org.apache.drill.exec.ops.BufferManagerImpl.getManagedBuffer(BufferManagerImpl.java:60) ~[drill-java-exec-1.6.0.jar:1.6.0]
	at org.apache.drill.exec.ops.BufferManagerImpl.getManagedBuffer(BufferManagerImpl.java:56) ~[drill-java-exec-1.6.0.jar:1.6.0]
	at org.apache.drill.exec.ops.QueryContext.getManagedBuffer(QueryContext.java:241) ~[drill-java-exec-1.6.0.jar:1.6.0]
	at org.apache.drill.exec.expr.fn.interpreter.InterpreterEvaluator$EvalVisitor.getManagedBufferIfAvailable(InterpreterEvaluator.java:158) ~[drill-java-exec-1.6.0.jar:1.6.0]

Cause:
Interpreter always asks for a new buffer to hold varchar/varbinary or decimal constant values. That's why the memory size required would be proportion to # of constant expressions multiplied by # of input rows (partition). This is different from evaluation from run-time generated where constant expression will be evaluated once and use only one buffer per value.

Fix:
To use one buffer for each unique constant value in query.

Issue Links

• links to: GitHub Pull Request #560