[ 
https://issues.apache.org/jira/browse/DRILL-7087?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17009342#comment-17009342
 ] 

Weijie Tong commented on DRILL-7087:
------------------------------------

Hi [~IhorHuzenko] , I have done this work for a long time. But since my job 
role has turned from Drill, I have no time to continue this work at the 
community. So hope someone could take this work, and sorry for that.

Your work is great. I think Drill should turn to Arrow to depend on a continual 
tuned memory format project. My solution is not a perfect final solution. We 
need to pay for the cost to count for the null tag bit to a nullable vector.
Completely turning Drill VV to Arrow is a large job to finish and review. Maybe 
an adapter to turn Drill's VV to Arrow's is needed. So we could finish this job 
one by one. But that's still a tough job, as you need to rewrite all the 
operators, storage plugins, and udfs.

> Integrate Arrow's Gandiva into Drill
> ------------------------------------
>
>                 Key: DRILL-7087
>                 URL: https://issues.apache.org/jira/browse/DRILL-7087
>             Project: Apache Drill
>          Issue Type: Improvement
>          Components: Execution - Codegen, Execution - Relational Operators
>            Reporter: Weijie Tong
>            Assignee: Weijie Tong
>            Priority: Major
>
> It's a prior work to integrate arrow into drill by invoking the its gandiva 
> feature. Comparing arrow and drill 's in memory column representation , 
> there's different null representation internal now. Drill use 1 byte while 
> arrow using 1 bit to indicate one null row. Also all columns of arrow is 
> nullable now. Apart from those basic differences , they have same memory 
> representation to the different data types. 
> The integrating strategy is to invoke arrow's JniWrapper's native method 
> directly by passing the ValueVector's memory address. 
> I have done a implementation at our own Drill version by integrating gandiva 
> into Drill's project operator. The performance shows that there's nearly 1 
> times performance gain at expression computation.
> So if there's no objection , I will submit a related PR to contribute this 
> feature. Also this issue waits for arrow's related issue[ARROW-4819].



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

Reply via email to