There are two kinds of intervals (in Calcite and in standard SQL). 
Days-hours-minutes-seconds-millseconds intervals, and year-month intervals. The 
former are represented internally in milliseconds. The latter are represented 
in months.

This is one area where Postgres does things differently from the SQL standard.


> On Feb 10, 2022, at 3:48 AM, Chathura Widanage <[email protected]> 
> wrote:
> 
> Thanks, Stamatis. Below is the original SQL query.
> 
> select l_returnflag, l_linestatus, sum(l_quantity) as sum_qty,
> sum(l_extendedprice) as sum_base_price, sum(l_extendedprice*(1-l_discount))
> as sum_disc_price, sum(l_extendedprice*(1-l_discount)*(1+l_tax)) as
> sum_charge, avg(l_quantity) as avg_qty, avg(l_extendedprice) as avg_price,
> avg(l_discount) as avg_disc, count(*) as count_order from lineitem where
> l_shipdate <= date '1998-12-01' - interval '90' day group by l_returnflag,
> l_linestatus order by l_returnflag, l_linestatus
> 
> Based on the shared line of code, even a month should be represented in
> milis right? But when the below query is transformed it shows months in
> months.
> 
> with revenue (suplier_no, total_revenue) as ( select l_suppkey,
> sum(l_extendedprice * (1-l_discount)) from lineitem where l_shipdate >=
> date '1996-01-01' and l_shipdate < date '1996-01-01' + interval '3' month
> group by l_suppkey ) select s_suppkey, s_name, s_address, s_phone,
> total_revenue from supplier, revenue where s_suppkey = suplier_no and
> total_revenue = ( select max(total_revenue) from revenue ) order by
> s_suppkey
> 
> AND(>=($1, 1993-07-01 00:00:00), <($1, CAST(+(1993-07-01, 3:INTERVAL
> MONTH)):TIMESTAMP(0) NOT NULL))
> 
> Thanks for the pointer Stamatis, I'll see whether there is something to do
> with the RexSimplify/RexExecutor.
> 
> Regards,
> Chathura
> 
> 
> On Thu, Feb 10, 2022 at 5:07 PM Stamatis Zampetakis <[email protected]>
> wrote:
> 
>> 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