Push down filters below `Unnest` in sub queries about arrow-datafusion HOT 11 CLOSED

I confirm that 1b396c4 fixes our more involved case in practice.

from arrow-datafusion.

I think this is related #10660

from arrow-datafusion.

Thanks for the pointers @duongcongtoai
I think the recursive and max_depth options can be a special workaround, but introduce a good amount of complexity. Fixing the push down makes single unnest composable in general.

from arrow-datafusion.

alternative query

Yes. The problem is that one can't move the filter to the inner query (views) or one shouldn't have to with other common sql constructs (sub queries, ctes).

from arrow-datafusion.

I think we can rewrite Unnest plan in push down filter

from arrow-datafusion.

Thanks for the swift action.

I just found out this second, that the change introduces a new issue when filtering on an unnested struct column (Optimizer rule 'common_sub_expression_eliminate' failed). Sorry for being late.

I'll file a new issue. The reproducer is less trivial as here.

from arrow-datafusion.

I confirm that 1b396c4 fixes our more involved case in practice.

Thank you @jayzhan211 for the quick response

from arrow-datafusion.

For this kind of query select unnest(unnest([[1],[1,2],[1,2,3]]));, I think we can follow what Duckdb does.

Unnest with recursive, and max_depth

select unnest([[1],[1,2],[1,2,3]], recursive := true);
or select unnest([[1],[1,2],[1,2,3]], max_depth := 2);

from arrow-datafusion.

@ahirner
I found an alternative query is to move filter into the query

select "unnest(v.column2)" from (select unnest(column2), column1 from v where column1=2);

query TT
explain select "unnest(v.column2)" from (select unnest(column2), column1 from v where column1=2);
----
logical_plan
01)Projection: unnest(v.column2)
02)--Unnest: lists[unnest(v.column2)] structs[]
03)----Projection: v.column2 AS unnest(v.column2), v.column1
04)------Filter: v.column1 = Int64(2)
05)--------TableScan: v projection=[column1, column2]
physical_plan
01)ProjectionExec: expr=[unnest(v.column2)@0 as unnest(v.column2)]
02)--UnnestExec
03)----RepartitionExec: partitioning=RoundRobinBatch(4), input_partitions=1
04)------ProjectionExec: expr=[column2@1 as unnest(v.column2), column1@0 as column1]
05)--------CoalesceBatchesExec: target_batch_size=8192
06)----------FilterExec: column1@0 = 2
07)------------MemoryExec: partitions=1, partition_sizes=[1]

Given that, it might not be necessary to support pushing down outer filter

from arrow-datafusion.

PR to do so: #10974

from arrow-datafusion.

Added issue that's quite minimal but maybe not entirely minimal: #10990

from arrow-datafusion.