[native] Push down FilterNode into TableScan

[Yuhta]

Sometimes the planner put a FilterNode right after TableScanNode. This change merges these 2 nodes by moving the filter expression into remaining filter of the table scan when possible. This results in fewer IO and CPU because table scan can leverage the information in remaining filter to skip file stripes in case of random sampling.

Also add $path and $bucket in split info columns and fix split counts in coordinator UI.

[aditi-pandit]

Thanks @Yuhta. Had a high level question:

Am curious why we are doing the pushdown of FilterNode into TableScan during Presto -> Velox plan conversion ? This could be an optimization in Velox Local Planner as well. That would make it common to Presto and Spark.

[Yuhta]

@aditi-pandit Local planner has no dependency on Hive so it does not have knowledge about remaining filter. I guess this was the same reason why Presto planner does not put it in remaining filter as well (did not verify). Seems Prestissimo is a reasonable place to fix this, as we have all the dependency needed and we have the knowledge that this will benefit the execution layer (Velox).

[aditi-pandit]

@aditi-pandit Local planner has no dependency on Hive so it does not have knowledge about remaining filter. I guess this was the same reason why Presto planner does not put it in remaining filter as well (did not verify). Seems Prestissimo is a reasonable place to fix this, as we have all the dependency needed and we have the knowledge that this will benefit the execution layer (Velox).

Hmm... Yeah, Velox Local Planner doesn't have dependency on Hive. Alright, your approach is reasonable.

[aditi-pandit]

Thanks @Yuhta for this code, and answering all my questions.

[jaystarshot]

Please consider adding release notes following our release notes guide - link.

Using below for now

*Merged FilterNode into TableScanNode where possible, reducing I/O and CPU; added $path and $bucket to split info, and fixed split counts in coordinator UI. :pr:23755``