Github user fhueske commented on a diff in the pull request:
https://github.com/apache/flink/pull/5043#discussion_r152577385
--- Diff:
flink-libraries/flink-table/src/main/scala/org/apache/flink/table/plan/nodes/logical/FlinkLogicalTableSourceScan.scala
---
@@ -62,7 +62,19 @@ class FlinkLogicalTableSourceScan(
override def computeSelfCost(planner: RelOptPlanner, metadata:
RelMetadataQuery): RelOptCost = {
val rowCnt = metadata.getRowCount(this)
- planner.getCostFactory.makeCost(rowCnt, rowCnt, rowCnt *
estimateRowSize(getRowType))
+
+ val adjustedCnt: Double = tableSource match {
+ case f: FilterableTableSource[_] if f.isFilterPushedDown =>
+ // ensure we prefer FilterableTableSources with pushed-down
filters.
+ rowCnt - 1.0
--- End diff --
Doesn't really make a difference IMO. It's all about relative costs. We
only need to make sure that a `FilterableTableSource` with pushed down
predicates appears to be less expensive than the same `TableSource` without
pushed predicates.
The problem has not occurred before, because the `OrcTableSource` does not
guarantee that all emitted rows match the pushed predicates. Therefore, all
predicates are also applied by a following `Calc` operator such that the cost
of that operator is not decreased.
---
