You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@calcite.apache.org by "Benchao Li (Jira)" <ji...@apache.org> on 2022/03/06 13:03:00 UTC
[jira] [Comment Edited] (CALCITE-5032) The unexpected behavior on method RelOptUtil#splitJoinCondition
[ https://issues.apache.org/jira/browse/CALCITE-5032?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17501698#comment-17501698 ]
Benchao Li edited comment on CALCITE-5032 at 3/6/22, 1:02 PM:
--------------------------------------------------------------
Hi [~luoyuxia] ,
I think you raised a valid question, the result for you case is not correct IMO.
For your case, I think there are two ways to handle it:
1. we require the condition to be simplified before call splitJoinCondition in the doc,
2. we fix it like you said: set the rightInput correctly.
was (Author: libenchao):
Hi [~luoyuxia] ,
I think you raised a valid question, the result for you case is not correct IMO.
For your case, I think there are two ways to handle it:
1. we require the condition to be simplified before call splitJoinCondition in the doc,
2. we fix it like you said: set the rightInput correctly.
However, I'm further confused about this part:
{code:java}
if ((rangeOp == null)
&& ((leftKey == null) || (rightKey == null))) {
// ...
} {code}
If we push this expr down into projection, the semantic of Join may be not correct, for example, `LEFT JOIN`.
> The unexpected behavior on method RelOptUtil#splitJoinCondition
> ---------------------------------------------------------------
>
> Key: CALCITE-5032
> URL: https://issues.apache.org/jira/browse/CALCITE-5032
> Project: Calcite
> Issue Type: Bug
> Components: core
> Reporter: luoyuxia
> Priority: Major
>
> When I call the method
> {code:java}
> RexNode splitJoinCondition(
> List<RelDataTypeField> sysFieldList,
> List<RelNode> inputs,
> RexNode condition,
> List<List<RexNode>> joinKeys,
> @Nullable List<Integer> filterNulls,
> @Nullable List<SqlOperator> rangeOp)
> {code}
> , and the passed `joinKeys` is list with size 2, `condition` is like "1 = 1", then I found the `joinKeys` passed dosen't change as expected.
> For the `joinKeys`, the first list will contain two RexNodes, and the second list contains no RexNode. I think the expected behavior should be the first list contain one RexNode, and the second contain one RexNode too.
> After I dive into the code, I found such code will be invoked:
> {code:java}
> if ((rangeOp == null)
> && ((leftKey == null) || (rightKey == null))) {
> // no equality join keys found yet:
> // try transforming the condition to
> // equality "join" conditions, e.g.
> // f(LHS) > 0 ===> ( f(LHS) > 0 ) = TRUE,
> // and make the RHS produce TRUE, but only if we're strictly
> // looking for equi-joins
> final ImmutableBitSet projRefs = InputFinder.bits(condition);
> leftKey = null;
> rightKey = null;
> boolean foundInput = false;
> for (int i = 0; i < inputs.size() && !foundInput; i++) {
> if (inputsRange[i].contains(projRefs)) {
> leftInput = i;
> leftFields = inputs.get(leftInput).getRowType().getFieldList();
> leftKey = condition.accept(
> new RelOptUtil.RexInputConverter(
> rexBuilder,
> leftFields,
> leftFields,
> adjustments));
> rightKey = rexBuilder.makeLiteral(true);
> // effectively performing an equality comparison
> kind = SqlKind.EQUALS;
> foundInput = true;
> }
> }
> }
> {code}
> It'll set `leftKey` and `righeKey`, but it won't set `rightInput`. So `rightInput` will be 0.
> Then it'll try to add `leftKey` and `righeKey` to `joinKeys`,
> {code:java}
> addJoinKey(
> joinKeys.get(leftInput),
> leftKey,
> (rangeOp != null) && !rangeOp.isEmpty());
> addJoinKey(
> joinKeys.get(rightInput),
> rightKey,
> (rangeOp != null) && !rangeOp.isEmpty());
> {code}
> But the `rightInput` is 0, so the first list of `joinKeys` will be also add `righeKey`.
> Could anyone please help me? Is it a bug or I misunderstand something? Thanks very much.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)