Implement native version of ColumnarToRow

[andygrove]

What is the problem the feature request solves?

Spark's ColumnarToRowExec can be very slow in some cases. This plan shows that C2R took 8 minutes even though the underlying scan only took 20 seconds. This example is selecting decimal columns, so this is related to #670 and #679, but implementing a native version of ColumnarToRowExec may help with this.

Describe the potential solution

No response

Additional context

No response

[andygrove]

I created a Google document to discuss the design.

https://docs.google.com/document/d/1zNuavf_WT3IcpeTVAEC8IjMGloi1MeAeQSg4F0eEivs/edit?usp=sharing

[andygrove]

I created #837 for the very first step in this process

[andygrove]

Here is a like for like comparison between Spark and Comet for Scan+C2R for the optimized version of q72.

Spark

Spark C2R takes 1.1 minutes

Comet

Comet C2R takes 1.3 minutes

[andygrove]

There is some spill happening, even though I am allocating 20g memory overhead:

[andygrove]

@parthchandra I know that you are working on this so I tried assigning the issue to you, but your name does not show up for me for some reason. I wondered if pinging you here would change that.

edit: No, that did not help

[parthchandra]

Posting a reply in case it helps associate the issue somehow. Anyhow, confirming that I am indeed working on this.
In that context, I am initially planning to only add support for fixed width primitive types. Then add decimals and varchar.
Estimating the sizes of complex types is somewhat complicated, so I will defer that until we see actual gains from this effort.