[
https://issues.apache.org/jira/browse/CALCITE-3865?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Haisheng Yuan updated CALCITE-3865:
-----------------------------------
Description:
CALCITE-1148 introduced the following change to RelCollationTraitDef to fix
RelTrait conversion bug, but it is just hiding the underlying issue and adding
redundant and unnecessary check to planner.
{code:java}
@Override public boolean canConvert(RelOptPlanner planner,
RelCollation fromTrait, RelCollation toTrait, RelNode fromRel) {
// Returns true only if we can convert. In this case, we can only convert
// if the fromTrait (the input) has fields that the toTrait wants to sort.
for (RelFieldCollation field : toTrait.getFieldCollations()) {
int index = field.getFieldIndex();
if (index >= fromRel.getRowType().getFieldCount()) {
return false;
}
}
return true;
}
{code}
The root cause is that logical operators, especially LogicalSort can have
traits, which is a bad design decision, and {{AggregateReduceFunctionsRule}}
fails to adjust the column mapping in RelTraitSet. The newly created
{{LogicalProject}} has collation on column 5 (it just copy its input's
RelTraitSet blindly), but it only has 2 columns.
!image-2020-03-19-16-44-25-621.png!
was:
CALCITE-1148 introduced the following change to RelCollationTraitDef to fix
RelTrait conversion bug, but it is just hiding the underlying issue and adding
redundant and unnecessary check to planner.
!image-2020-03-19-16-47-21-805.png!
The root cause is that logical operators, especially LogicalSort can have
traits, which is a bad design decision, and {{AggregateReduceFunctionsRule}}
fails to adjust the column mapping in RelTraitSet. The newly created
{{LogicalProject}} has collation on column 5 (it just copy its input's
RelTraitSet blindly), but it only has 2 columns.
!image-2020-03-19-16-44-25-621.png!
> RelCollationTraitDef.canConvert should always return true
> ---------------------------------------------------------
>
> Key: CALCITE-3865
> URL: https://issues.apache.org/jira/browse/CALCITE-3865
> Project: Calcite
> Issue Type: Improvement
> Components: core
> Reporter: Haisheng Yuan
> Priority: Major
> Attachments: image-2020-03-19-16-44-25-621.png,
> image-2020-03-19-16-47-21-805.png
>
>
> CALCITE-1148 introduced the following change to RelCollationTraitDef to fix
> RelTrait conversion bug, but it is just hiding the underlying issue and
> adding redundant and unnecessary check to planner.
> {code:java}
> @Override public boolean canConvert(RelOptPlanner planner,
> RelCollation fromTrait, RelCollation toTrait, RelNode fromRel) {
> // Returns true only if we can convert. In this case, we can only convert
> // if the fromTrait (the input) has fields that the toTrait wants to sort.
> for (RelFieldCollation field : toTrait.getFieldCollations()) {
> int index = field.getFieldIndex();
> if (index >= fromRel.getRowType().getFieldCount()) {
> return false;
> }
> }
> return true;
> }
> {code}
> The root cause is that logical operators, especially LogicalSort can have
> traits, which is a bad design decision, and {{AggregateReduceFunctionsRule}}
> fails to adjust the column mapping in RelTraitSet. The newly created
> {{LogicalProject}} has collation on column 5 (it just copy its input's
> RelTraitSet blindly), but it only has 2 columns.
> !image-2020-03-19-16-44-25-621.png!
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
