[
https://issues.apache.org/jira/browse/DRILL-6513?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16517641#comment-16517641
]
ASF GitHub Bot commented on DRILL-6513:
---------------------------------------
rhou1 commented on a change in pull request #1329: DRILL-6513: Max query memory
per node set to Drillbit's max direrct memory
URL: https://github.com/apache/drill/pull/1329#discussion_r196616430
##########
File path:
exec/java-exec/src/main/java/org/apache/drill/exec/ExecConstants.java
##########
@@ -440,7 +441,7 @@ private ExecConstants() {
* DEFAULT: 2048 MB
*/
public static final String MAX_QUERY_MEMORY_PER_NODE_KEY =
"planner.memory.max_query_memory_per_node";
- public static final LongValidator MAX_QUERY_MEMORY_PER_NODE = new
RangeLongValidator(MAX_QUERY_MEMORY_PER_NODE_KEY, 1024 * 1024, Long.MAX_VALUE);
+ public static final LongValidator MAX_QUERY_MEMORY_PER_NODE = new
RangeLongValidator(MAX_QUERY_MEMORY_PER_NODE_KEY, 1024 * 1024,
DrillConfig.getMaxDirectMemory());
Review comment:
The current implementation enables the customer to set
MAX_QUERY_MEMORY_PER_NODE to unreasonable values, such as 1 TB. This has lead
to out-of-memory issues in some of our testing.
Drill should validate user settings to a reasonable point. This is an
attempt to move in that direction. There is more work needed here.
I think it is challenging for the planner to make the best optimizations if
each node in the cluster is a different size, and they want to use each node to
its best capabilities. I.e. make some nodes work harder than others.
One work-around is for the customer to set this option on the node with the
largest Direct Memory if they want to set the option higher than recommended.
This could cause instability on the nodes that have less Direct Memory
available.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
> Drill should only allow valid values when users set
> planner.memory.max_query_memory_per_node
> --------------------------------------------------------------------------------------------
>
> Key: DRILL-6513
> URL: https://issues.apache.org/jira/browse/DRILL-6513
> Project: Apache Drill
> Issue Type: Bug
> Components: Query Planning & Optimization
> Reporter: salim achouche
> Assignee: salim achouche
> Priority: Major
> Labels: pull-request-available
> Fix For: 1.14.0
>
>
> The "planner.memory.max_query_memory_per_node" configuration can be currently
> set to values higher than the Drillbit Direct Memory configuration. The goal
> of this Jira is to fail queries with such an erroneous configuration to avoid
> runtime OOM.
> NOTE - The current semantic of the maximum query memory per node
> configuration is that the end user has computed valid values especially
> knowing the current Drill limitations. Such values have to account for
> Netty's overhead (memory pools), shared pools (e.g., network exchanges), and
> concurrent query execution. This Jira should not be used to also cover such
> use-cases. The Drill Resource Management feature has the means to automate
> query quotas and the associated validation. We should create another Jira
> requesting the enhanced validations contracts under the umbrella of the
> Resource Management feature.
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
