[
https://issues.apache.org/jira/browse/CALCITE-5479?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17676980#comment-17676980
]
Gian Merlino commented on CALCITE-5479:
---------------------------------------
Ah. Would you mind explaining what the behavior is supposed to be? Maybe I
don't understand how {{OperandTypes.sequence}} is supposed to work. The stuff
we're doing in Druid that stopped working is like:
{code}
OperandTypes.sequence(
"F(expr, path)",
OperandTypes.ANY,
OperandTypes.and(OperandTypes.family(SqlTypeFamily.STRING),
OperandTypes.LITERAL)
)
{code}
That's for an operator that accepts an {{expr}}, which can be any type, and a
{{path}}, which must be a string literal.
The sequence checker passes {{iFormalOperand = 1}} to the {{and}} checker. Then
the checker {{OperandTypes.family(SqlTypeFamily.STRING)}} throws an error on
{{families.get(iFormalOperand)}}, because {{iFormalOperand}} is {{1}} but
{{families}} doesn't have that many elements.
If we're using the APIs wrong, please let me know since I'd rather fix that on
our end.
> Restore consistent handling of iFormalOperand in sequence checkers
> ------------------------------------------------------------------
>
> Key: CALCITE-5479
> URL: https://issues.apache.org/jira/browse/CALCITE-5479
> Project: Calcite
> Issue Type: Bug
> Reporter: Gian Merlino
> Priority: Major
> Labels: pull-request-available
> Time Spent: 20m
> Remaining Estimate: 0h
>
> Handling for {{OperandTypes.sequence}} changed in
> [33f4ab40bbee26e06209061c35a422f2f1e05371|https://github.com/apache/calcite/commit/33f4ab40bbee26e06209061c35a422f2f1e05371#diff-b0b8d58a792b8e60b9e97717912aecfc6695536f5026ac4d5231d14e34b91566L303-R316]
> such that {{iFormalOperand}} passed to subcheckers is no longer always zero,
> but is instead:
> - Zero if the subchecker is {{FamilyOperandTypeChecker}}.
> - Otherwise, the operand number in the overall sequence.
> It causes problems for the way we're using sequence checkers in Druid, since
> we don't always use {{FamilyOperandTypeChecker}}, but we _do_ assume the old
> behavior: that {{iFormalOperand}} is always zero, and therefore we can put
> any checker into the sequence without it being "aware" that it is in a
> sequence.
> I marked this as a bug in case this change was made accidentally. If it was
> made for a reason, please let me know. Thanks.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
