Thomas Rebele created CALCITE-5473:
--------------------------------------
Summary: Keep more collations in RelMdCollation#project for
monotonic RexCalls
Key: CALCITE-5473
URL: https://issues.apache.org/jira/browse/CALCITE-5473
Project: Calcite
Issue Type: Improvement
Reporter: Thomas Rebele
Some RexCalls do not affect the collation, i.e., $0+10 has the same collation
as $0 (if there are now integer overflows). This is already implemented in
{{{}SqlMonotonicBinaryOperator#getMonotonicity{}}}.
If we have an input with collation [0,1], and a {{{}Calc(a=$0+10, b=$1){}}},
then the output collation would be [0, 1] as well. This is the case for
STRICTLY_INCREASING or STRICTLY_DECREASING calls (though the direction of the
field collation might need to be adapted).
However, if the Calc was {{{}Calc(a=FLOOR($0), b=$1){}}}, then the output
collation would be just [0], as the sortedness of b is not guaranteed. This
happens for example with rows (0.2, 50) and (0.3, 20) and (0.4, 30), in that
order. So for INCREASING, DECREASING, or CONSTANT the collation would be
shortened.
The right place to implement this would probably be
{{{}RelMdCollation#project{}}}. The {{fieldCollationsForRexCalls}} loop would
need to stay, because some calls may introduce a collation, regardless of the
input. E.g., {{SELECT CURRENT_DATE FROM some_random_table}} is always sorted.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
