Hi Chathura,

It is difficult to reason about correctness without having the actual SQL
query at hand.

The fact that you have milliseconds is not by itself a problem and has to
do with the way Calcite internally represents intervals (see comment in
[1]).

Also from the examples you provided the behavior in 1.29.0 does not seem to
be an additional transformation rather than a missing simplification
(constant reduction). I am not sure if this is intentional or not but I
guess you can have a look at the changes landed around
RexSimplify/RexExecutor.

Best,
Stamatis

[1]
https://github.com/apache/calcite/blob/812e3e98eae518cf85cd1b6b7f055fb96784a423/core/src/main/java/org/apache/calcite/rex/RexLiteral.java#L357


On Thu, Feb 10, 2022 at 8:02 AM Chathura Widanage <
[email protected]> wrote:

> Hi community,
>
> I'm comparing two rel expressions generated by calcite 1.25.0 and 1.29.0
> and noticed there is an invalid IntervalSQLType plugged into the query.
>
> <=($6, 1998-09-02 00:00:00) : Calcite 1.25.0
> vs
> <=($6, CAST(-(1998-12-01, 7776000000:INTERVAL DAY)):TIMESTAMP(0) NOT NULL)
> :
> Calcite 1.29.0
>
> 7776000000 is 90 days in milliseconds, but the IntervalSQLType/value
> combination is invalid.
>
> Could you please let me know whether this could be a bug and whether there
> an option to prevent such transformations at all?
>
> Regards,
> Chathura
>
> PS: This comes on queries from tpch benchmark and invalid conversion is
> from tpch-01.
>
> I'm seeing similar conversions in other queries, but they seem to be
> correct, but feels this transformation is redundant.
>
> AND(>=($1, 1993-07-01 00:00:00), <($1, 1993-10-01 00:00:00))
> vs
> AND(>=($1, 1993-07-01 00:00:00), <($1, CAST(+(1993-07-01, 3:INTERVAL
> MONTH)):TIMESTAMP(0) NOT NULL))
>

Reply via email to