[
https://issues.apache.org/jira/browse/CALCITE-3711?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17008696#comment-17008696
]
Ruben Q L commented on CALCITE-3711:
------------------------------------
[~vladimirsitnikov], that was my first intention too, but the problem is that
Correlate is not a Join (it extends directly from BiRel) and it does not have
the notion of a RexNode "condition" (as Join does), so I would say it is not
possible to apply {{RelMdUtil#getJoinRowCount}} approach (which is based on
condition's selectivity) into a Correlate.
Moreover, even though one might usually think of Correlate as a synonym of
"Correlated Join", this is not always the case, we may have Correlate operators
(LogicalCorrelate and EnumerableCorrelate) to implement operations other than
joins (e.g. an UNNEST).
> Correlate should override estimateRowCount
> ------------------------------------------
>
> Key: CALCITE-3711
> URL: https://issues.apache.org/jira/browse/CALCITE-3711
> Project: Calcite
> Issue Type: Improvement
> Affects Versions: 1.21.0
> Reporter: Ruben Q L
> Priority: Major
>
> Currently, Correlate inherits the {{estimateRowCount}} implementation from
> AbstractRelNode:
> {code:java}
> public double estimateRowCount(RelMetadataQuery mq) {
> return 1.0;
> }
> {code}
> which gives a "default" (but unrealistic) row count computation.
> This has several issues:
> - This value may be used by {{Correlate#computeSelfCost}} (and by inheritance
> by LogicalCorrelate / EnumerableCorrelate):
> {code:java}
> @Override public RelOptCost computeSelfCost(final RelOptPlanner planner,
> final RelMetadataQuery mq) {
> double rowCount = mq.getRowCount(this); // by default: estimateRowCount,
> i.e. 1.0
> ...
> {code}
> - As the current state, this value will definitely by used by several join
> algorithms in their cost computation if their left/right child is a Correlate:
> {code:java}
> public class EnumerableHashJoin extends Join implements EnumerableRel {
> ...
> @Override public RelOptCost computeSelfCost(RelOptPlanner planner,
> RelMetadataQuery mq) {
> ...
> final double rightRowCount = right.estimateRowCount(mq); // 1.0 if right
> is a Correlate
> final double leftRowCount = left.estimateRowCount(mq); // 1.0 if left is
> a Correlate
> ...
> {code}
> Even though cost computation is pluggable (which would help solving issues
> like the first one, but not issues like the second one), IMHO we should
> provide a more realistic default {{Correlate#estimateRowCount}} computation.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
