I agree with your observation that if (x, y) is a unique key for R, and x is
constant, then y is also a unique key.
When someone calls areColumnsUnique(R, {y}, false), the implementation should
(and probably already does) look to see whether {y} is a known unique key (or a
superset of one).
But the implementation of getUniqueKeys(R) probably doesn’t yet take into
account constant columns. If the known unique keys are {x, y}, {w, x, z}, {y,
z}, and {x} is constant, then a better set of unique keys is {y}, {w, z}. (Note
that we have converted {x, y} to {y}, and we have discarded {y, z} because it
is a superset of the new key.)
That would be a good enhancement to make to getUniqueKeys. Note that it doesn’t
depend on the type of the RelNode R; it just relies on the constants known for
R, via RelOptPredicateList.constantMap.
Some work seems to have been done on this in
https://issues.apache.org/jira/browse/CALCITE-3428.
Julian
> On Oct 25, 2023, at 12:35 PM, Paul Jackson
> <[email protected]> wrote:
>
> I'm currently working on CALCITE-6044. I have a problem that I thought would
> be solved by mq.getAllPredicates(rel) but the output is from a perspective
> that doesn't seem useful for my problem, which is, given this query:
>
> select deptno, ename, sum(sal)from empwhere deptno=1010group by deptno, ename
> and a call to areColumnsUnique(Aggregate rel, mq, bitSetOf(1), false), I'd
> like it to return true. It currently returns true for columns {0, 1} on
> account that those are the columns in the group by clause. However, it
> doesn't recognize that since deptno is set to a constant, the remaining
> column (ename/1) is by itself an effective key.
> To solve this, I plan to collect a bitset of the constant columns (would be
> {0} in this example), and remove those columns from the group-by columns. I
> thought the way to collect those columns was to call mq.getAllPredicates(rel)
> and take the columns from constantMap.keySet(), however, the RexInputRefs are
> relative to the table scans (returning column 7 instead of 0), ignoring the
> effects on column ordering of the projection and aggregation.
> Is there an existing utility that can either transform these results or
> return them from the perspective of the current RelNode?
> Paul Jackson
