drill-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Charles Givre <cgi...@gmail.com>
Subject Excessive Memory Use in Parquet Files (From Drill Slack Channel)
Date Tue, 24 Mar 2020 18:46:40 GMT

Idan Sheinberg  8:21 AM
Hi  there
I'm trying run a simple offset query (ORDER BY timestamp LIMIT 500 OFFSET 1000) against rather
complex parquet files (say 4 columns, once being an array currently consisting of a single
element comprised of 15 columns)
All files share the same Schema, of course.
 User Error Occurred: One or more nodes ran out of memory while executing the query. (null)
org.apache.drill.common.exceptions.UserException: RESOURCE ERROR: One or more nodes ran out
of memory while executing the query.
[Error Id: 67b61fc9-320f-47a1-8718-813843a10ecc ]
	at org.apache.drill.common.exceptions.UserException$Builder.build(UserException.java:657)
	at org.apache.drill.exec.work.fragment.FragmentExecutor.run(FragmentExecutor.java:338)
	at org.apache.drill.common.SelfCleaningRunnable.run(SelfCleaningRunnable.java:38)
	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
	at java.lang.Thread.run(Thread.java:748)
Caused by: org.apache.drill.exec.exception.OutOfMemoryException: null
	at org.apache.drill.exec.vector.complex.AbstractContainerVector.allocateNew(AbstractContainerVector.java:59)
	at org.apache.drill.exec.test.generated.PartitionerGen5$OutgoingRecordBatch.allocateOutgoingRecordBatch(PartitionerTemplate.java:380)
	at org.apache.drill.exec.test.generated.PartitionerGen5$OutgoingRecordBatch.initializeBatch(PartitionerTemplate.java:400)
	at org.apache.drill.exec.test.generated.PartitionerGen5.setup(PartitionerTemplate.java:126)
	at org.apache.drill.exec.physical.impl.partitionsender.PartitionSenderRootExec.createClassInstances(PartitionSenderRootExec.java:263)
	at org.apache.drill.exec.physical.impl.partitionsender.PartitionSenderRootExec.createPartitioner(PartitionSenderRootExec.java:218)
	at org.apache.drill.exec.physical.impl.partitionsender.PartitionSenderRootExec.innerNext(PartitionSenderRootExec.java:188)
	at org.apache.drill.exec.physical.impl.BaseRootExec.next(BaseRootExec.java:93)
	at org.apache.drill.exec.work.fragment.FragmentExecutor$1.run(FragmentExecutor.java:323)
	at org.apache.drill.exec.work.fragment.FragmentExecutor$1.run(FragmentExecutor.java:310)
	at java.security.AccessController.doPrivileged(Native Method)
	at javax.security.auth.Subject.doAs(Subject.java:422)
	at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1730)
	at org.apache.drill.exec.work.fragment.FragmentExecutor.run(FragmentExecutor.java:310)
	... 4 common frames omitted
Now, I'm running this query from a 16 core, 32GB Ram machine, with Heap sized at 20GB, Eden
sized at 16GB (added manually to JAVA_OPTS) and Direct Sized at 8 GB.
By querying sys.memory I can confirm all limits apply. At no point throughout the query Am
I nearing memory limit of the HEAP/DIRECT or the OS itself

However, due to the way org.apache.drill.exec.vector.complex.AbstractContainerVector.allocateNew
is impelmented
  public void allocateNew() throws OutOfMemoryException {
    if (!allocateNewSafe()) {
      throw new OutOfMemoryException();
The actual exception/error is swallowed, and I have no idea what's the cause of the failure
The data-set itself consists of say 15 parquet files, each one weighing at about 100kb
but as mentioned earlier, the parquet files are a bit more complex than the usual.
@cgivre @Vova Vysotskyi is there anything I can do or tweak to make this error go away?

cgivre  8:40 AM
This may be a bug.  Can you create an issue on our JIRA board?

Idan Sheinberg  8:43 AM
I'll get to it

cgivre  8:44 AM
I'd like for Paul Rogers to see this as I think he was the author of some of this.

Idan Sheinberg  8:44 AM
Hmm. I'll keep that in mind

cgivre  8:47 AM
We've been refactoring some of the complex readers as well, so its possible that is caused
this, but I'm not really sure.
What version of Drill?

cgivre  9:11 AM
This kind of info is super helpful as we're trying to work out all these details.
Reading schemas on the fly is not trivial, so when we find issues, we do like to resolve them

Idan Sheinberg  9:16 AM
This is drill 0.18 -SNAPSHOT as of last month
I do think I managed to resolve the issue however
I'm going to run some additional tests and let you know

cgivre  9:16 AM
What did you do?
You might want to rebase with today's build as well

Idan Sheinberg  9:21 AM
I'll come back with the details in a few moments

cgivre  9:38 AM
new messages

Idan Sheinberg  9:50 AM
Ok. See it seems as though it's a combination of a few things.
The data-set in question is still small (as mentioned before), but we are setting planner.slice_target
 to an extremely low value in order to trigger parallelism and speed up parquet parsing by
using multiple fragments.
We have 16 cores, 32 GB (C5.4xlarge on AWS) but we set planner.width.max_per_node  to further
increase parallelism.  it seems as though each fragment is handling parquet parsing on it's
own, and somehow incurs a great burden on
the direct memory buffer pool, as I do see 16GB peaks of direct memory usage after lowering
the planner.width.max_per_node to 16 (our available core).
The query planner itself reports the HASH_PARTITION_SENDER as the largest phase with 1-2 GB
of memory utilization
Seeing such an impact (16 GB of direct memory) for 1K items spread across 15 files, even with
a very complex parquet schema, seems unreasonable to me
View raw message