drill-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "ASF GitHub Bot (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (DRILL-6071) Limit batch size for flatten operator
Date Thu, 18 Jan 2018 01:23:02 GMT

    [ https://issues.apache.org/jira/browse/DRILL-6071?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16329835#comment-16329835
] 

ASF GitHub Bot commented on DRILL-6071:
---------------------------------------

Github user paul-rogers commented on a diff in the pull request:

    https://github.com/apache/drill/pull/1091#discussion_r162225383
  
    --- Diff: exec/java-exec/src/main/java/org/apache/drill/exec/physical/base/AbstractBase.java
---
    @@ -29,9 +28,12 @@
     
       public static long INIT_ALLOCATION = 1_000_000L;
       public static long MAX_ALLOCATION = 10_000_000_000L;
    +  // Default output batch size, 512MB
    +  public static long OUTPUT_BATCH_SIZE = 512 * 1024 * 1024L;
    --- End diff --
    
    Too large. The sort & hash agg operators often receive just 20-40 MB on a large cluster.
(That is, itself, an issue, but one that has proven very difficult to resolve.) So, the output
batch size must be no larger than 1/3 this size (for sort). Probably some team discussion
is required to agree on a good number, and on the work needed to ensure that sort, hash agg
and hash join are given sufficient memory for the selected batch size.


> Limit batch size for flatten operator
> -------------------------------------
>
>                 Key: DRILL-6071
>                 URL: https://issues.apache.org/jira/browse/DRILL-6071
>             Project: Apache Drill
>          Issue Type: Bug
>          Components: Execution - Flow
>    Affects Versions: 1.12.0
>            Reporter: Padma Penumarthy
>            Assignee: Padma Penumarthy
>            Priority: Major
>             Fix For: 1.13.0
>
>
> flatten currently uses an adaptive algorithm to control the outgoing batch size. 
> While processing the input batch, it adjusts the number of records in outgoing batch
based on memory usage so far. Once memory usage exceeds the configured limit for a batch,
the algorithm becomes more proactive and adjusts the limit half way through  and end of every
batch. All this periodic checking of memory usage is unnecessary overhead and impacts performance.
Also, we will know only after the fact. 
> Instead, figure out how many rows should be there in the outgoing batch from incoming
batch.
> The way to do that would be to figure out average row size of the outgoing batch and
based on that figure out how many rows can be there for a given amount of memory. value vectors
provide us the necessary information to be able to figure this out.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Mime
View raw message