[jira] [Commented] (DRILL-3913) Possible memory leak during CTAS using 30 TB TPC-H dataset

["Victoria Markman (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/DRILL-3913?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14947825#comment-14947825 ] 

Victoria Markman commented on DRILL-3913:
-----------------------------------------

[~agirish] 
1. Can you please run one single CTAS query in a loop and see if it fails with OOM at some point ? 
    Simple script is attached in: https://issues.apache.org/jira/browse/DRILL-2865
2. I believe reproduction can be scaled down: there is no need to run with 30TB of data
3. It is also possible, that you are running into an issue with memory fragmentation that [~adeneche] and I were trying to get to the bottom of couple of weeks ago, but got distracted by 1.2 release and never finished.

I think this is very important to understand what the source of the problem is in this case, raising priority to critical.

> Possible memory leak during CTAS using 30 TB TPC-H dataset
> ----------------------------------------------------------
>
>                 Key: DRILL-3913
>                 URL: https://issues.apache.org/jira/browse/DRILL-3913
>             Project: Apache Drill
>          Issue Type: Bug
>          Components: Execution - Flow
>    Affects Versions: 1.2.0
>         Environment: 47 nodes configured with 32 GB Drill Direct memory
>            Reporter: Abhishek Girish
>         Attachments: create_table_sf30000.txt, query_profile.json, sys.memory.txt
>
>
> 8 CTAS queries were executed sequentially to write TPC-H text data into Parquet. After successfully writing a few tables, CTAS failed with OOM.
> Restarting Drillbits fixed the problem and re-run of pending CTAS queries completed. This process was done twice in-order to complete all 8 tables to be written. Overall source was 30TB in size. 
> Queries attached. Query profile for one of the CTAS which failed is attached. Logs indicated that the Drillbit was out of Direct Memory. 
> Can share more details as required.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)