Ruben Q L created CALCITE-4995:
----------------------------------
Summary: ArrayIndexOutOfBoundsException caused by RelFieldTrimmer
on SEMI/ANTI join
Key: CALCITE-4995
URL: https://issues.apache.org/jira/browse/CALCITE-4995
Project: Calcite
Issue Type: Bug
Components: core
Affects Versions: 1.29.0
Reporter: Ruben Q L
Assignee: Ruben Q L
(Unit test to be provided)
It seems {{RelFieldTrimmer}} can cause an {{ArrayIndexOutOfBoundsException}} on
certain plans involving SEMI/ANTI join (i.e. joins that do NOT project the RHS
fields).
The root cause seems to be the "early return" in
{{RelFieldTrimmer#trimFields(Join join, ImmutableBitSet fieldsUsed,
Set<RelDataTypeField> extraFields)}} when nothing has been trimmed inside
join's inputs (so the join itself can be return as it is):
{code:java}
if (changeCount == 0
&& mapping.isIdentity()) {
return result(join, Mappings.createIdentity(fieldCount));
}
{code}
The problem is that this {{fieldCount}} is an addition of LHS + RHS fields (+
system fields); but in case of a SEMI/ANTI the mappings to be returned must not
consider RHS fields (since they are not projected by these join types).
The problem only happens here (when the trimmer does not trim the join). Notice
that, a few lines below, in the "other return scenario" of the method (when
something has been trimmed), there is a special treatment of the mapping for
ANTI/SEMI, so things will work fine in this case:
{code:java}
switch (join.getJoinType()) {
case SEMI:
case ANTI:
// For SemiJoins and AntiJoins only map fields from the left-side
if (join.getJoinType() == JoinRelType.SEMI) {
relBuilder.semiJoin(newConditionExpr);
} else {
relBuilder.antiJoin(newConditionExpr);
}
Mapping inputMapping = inputMappings.get(0);
mapping = Mappings.create(MappingType.INVERSE_SURJECTION,
join.getRowType().getFieldCount(),
newSystemFieldCount + inputMapping.getTargetCount());
for (int i = 0; i < newSystemFieldCount; ++i) {
mapping.set(i, i);
}
offset = systemFieldCount;
newOffset = newSystemFieldCount;
for (IntPair pair : inputMapping) {
mapping.set(pair.source + offset, pair.target + newOffset);
}
break;
default:
relBuilder.join(join.getJoinType(), newConditionExpr);
}
relBuilder.hints(join.getHints());
return result(relBuilder.build(), mapping);
{code}
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
