Andrew Lamb created ARROW-9771:
----------------------------------
Summary: [Rust] [DataFusion] Predicate Pushdown Improvement: treat
predicates separated by AND separately
Key: ARROW-9771
URL: https://issues.apache.org/jira/browse/ARROW-9771
Project: Apache Arrow
Issue Type: Improvement
Reporter: Andrew Lamb
As discussed by [~jorgecarleitao] and [~houqp] here:
https://github.com/apache/arrow/pull/7880#pullrequestreview-468057624
If a predicate is a conjunction (aka AND'd) together, each of the clauses can
be treated separately (e.g. a single filter expression {{A > 5 And B < 4}} can
be broken up and each of {{A > 5}} and {{B < 4}} can be potentially pushed down
different levels
The filter pushdown logic works for the following case (when {{a}} and {{b}}
are are separate selections, predicate for a is pushed below the {{Aggregate}}
in the optimized plan):
{code}
********Original plan:
Selection: #b GtEq Int64(1)
Selection: #a LtEq Int64(1)
Aggregate: groupBy=[[#a]], aggr=[[MIN(#b)]]
TableScan: test projection=None
********Optimized plan:
Selection: #b GtEq Int64(1)
Aggregate: groupBy=[[#a]], aggr=[[MIN(#b)]]
Selection: #a LtEq Int64(1)
TableScan: test projection=None
{code}
But not for this when {{a}} and {{b}} are {{AND}}'d together
{code}
********Original plan:
Selection: #a LtEq Int64(1) And #b GtEq Int64(1)
Aggregate: groupBy=[[#a]], aggr=[[MIN(#b)]]
TableScan: test projection=None
********Optimized plan:
Selection: #a LtEq Int64(1) And #b GtEq Int64(1)
Aggregate: groupBy=[[#a]], aggr=[[MIN(#b)]]
TableScan: test projection=None
{code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
