[
https://issues.apache.org/jira/browse/CALCITE-2898?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16786514#comment-16786514
]
Ruben Quesada Lopez edited comment on CALCITE-2898 at 3/7/19 8:34 AM:
----------------------------------------------------------------------
To provide a bit of context, I got into this issue while creating a semijoin
involving a condition "field = expression.field". In spite of the problem
described in this ticket (condition not identified as equijoin), the semijoin
is created fine by SemiJoinFactoryImpl:
{code:java}
private static class SemiJoinFactoryImpl implements SemiJoinFactory {
public RelNode createSemiJoin(RelNode left, RelNode right,
RexNode condition) {
final JoinInfo joinInfo = JoinInfo.of(left, right, condition);
return SemiJoin.create(left, right,
condition, joinInfo.leftKeys, joinInfo.rightKeys);
}
}
{code}
However, during the plan computation, SemiJoin#copy method is executed, which
contains an assertion to verify that we are dealing with an equijoin:
{code:java}
@Override public SemiJoin copy(RelTraitSet traitSet, RexNode condition,
RelNode left, RelNode right, JoinRelType joinType, boolean semiJoinDone) {
assert joinType == JoinRelType.INNER;
final JoinInfo joinInfo = JoinInfo.of(left, right, condition);
assert joinInfo.isEqui(); // !!! <= AssertionError
return new SemiJoin(getCluster(), traitSet, left, right, condition,
joinInfo.leftKeys, joinInfo.rightKeys);
}
{code}
In my case, I got an AssertionError (joinInfo is not equijoin), but this seems
to be a "false" error, because if I run my test with assertions disabled, it
runs successfully: my semijoin with the condition $0 = $1.id returns the
expected results.
Therefore, I believe either the assertion should be removed (if there is no
assertion when we create the semijoin, I don't see why there should be one when
copying it); or conditions like the one on my test should be detected as
equijoin (as proposed in the PR).
was (Author: rubenql):
To provide a bit of context, I got into this issue while creating a semijoin
involving an expression "field = expression.field". In spite of the problem
described in this ticket (condition not idenfitifed as equijoin), the semijoin
is created fine by SemiJoinFactoryImpl:
{code:java}
private static class SemiJoinFactoryImpl implements SemiJoinFactory {
public RelNode createSemiJoin(RelNode left, RelNode right,
RexNode condition) {
final JoinInfo joinInfo = JoinInfo.of(left, right, condition);
return SemiJoin.create(left, right,
condition, joinInfo.leftKeys, joinInfo.rightKeys);
}
}
{code}
However, during the plan computation, SemiJoin#copy method is executed, which
contains an assertion to verify that we are dealing with an equijoin:
{code:java}
@Override public SemiJoin copy(RelTraitSet traitSet, RexNode condition,
RelNode left, RelNode right, JoinRelType joinType, boolean semiJoinDone) {
assert joinType == JoinRelType.INNER;
final JoinInfo joinInfo = JoinInfo.of(left, right, condition);
assert joinInfo.isEqui(); // !!! <= AssertionError
return new SemiJoin(getCluster(), traitSet, left, right, condition,
joinInfo.leftKeys, joinInfo.rightKeys);
}
{code}
In my case, I got an AssertionError (joinInfo is not equijoin), but this seems
to be a "false" error, because if I run my test with assertions disabled, it
runs successfully. So, I believe either the assertion should be removed (if
there is no assertion when we create the semijoin, I don't see why there should
be one when copying it); or conditions like the one on my tests should be
detected as equijoin (as proposed in the PR).
> RelOptUtil#splitJoinCondition must consider RexFieldAccess referencing
> RexInputRef
> ----------------------------------------------------------------------------------
>
> Key: CALCITE-2898
> URL: https://issues.apache.org/jira/browse/CALCITE-2898
> Project: Calcite
> Issue Type: Bug
> Affects Versions: 1.18.0
> Reporter: Ruben Quesada Lopez
> Assignee: Ruben Quesada Lopez
> Priority: Major
> Labels: pull-request-available
> Time Spent: 10m
> Remaining Estimate: 0h
>
> {{RelOptUtil#splitJoinCondition}} "Splits out the equi-join components of a
> join condition, and returns what's left (remaining join filters that are not
> equijoins)". This works fine in case of RexInputRef operands in the condition
> (e.g. $0 = $1), but if any of the operands is a RexFieldAccess referencing a
> RexInputRef (e.g. $0 = $1.id), then the condition will NOT be detected as an
> equi-join and will be returned as if it were a non-equijoin.
> This can lead to undesired consequences, e.g {{JoinInfo#of}} would return a
> NonEquiJoinInfo object instead of an EquiJoinInfo object, which can generate
> problems if, for example, we are creating a SemiJoin (which requires an
> EquiJoinInfo object)
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
