[SPARK-49261][SQL] Don't replace literals in aggregate expressions with group-by expressions

[bersprockets]

What changes were proposed in this pull request?

Before this PR, RewriteDistinctAggregates could potentially replace literals in the aggregate expressions with output attributes from the Expand operator. This can occur when a group-by expression is a literal that happens by chance to match a literal used in an aggregate expression. E.g.:

create or replace temp view v1(a, b, c) as values
(1, 1.001d, 2), (2, 3.001d, 4), (2, 3.001, 4);

cache table v1;

select
  round(sum(b), 6) as sum1,
  count(distinct a) as count1,
  count(distinct c) as count2
from (
  select
    6 as gb,
    *
  from v1
)
group by a, gb;

In the optimized plan, you can see that the literal 6 in the round function invocation has been patched with an output attribute (6#163) from the Expand operator:

== Optimized Logical Plan ==
'Aggregate [a#123, 6#163], [round(first(sum(__auto_generated_subquery_name.b)#167, true) FILTER (WHERE (gid#162 = 0)), 6#163) AS sum1#114, count(__auto_generated_subquery_name.a#164) FILTER (WHERE (gid#162 = 1)) AS count1#115L, count(__auto_generated_subquery_name.c#165) FILTER (WHERE (gid#162 = 2)) AS count2#116L]
+- Aggregate [a#123, 6#163, __auto_generated_subquery_name.a#164, __auto_generated_subquery_name.c#165, gid#162], [a#123, 6#163, __auto_generated_subquery_name.a#164, __auto_generated_subquery_name.c#165, gid#162, sum(__auto_generated_subquery_name.b#166) AS sum(__auto_generated_subquery_name.b)#167]
   +- Expand [[a#123, 6, null, null, 0, b#124], [a#123, 6, a#123, null, 1, null], [a#123, 6, null, c#125, 2, null]], [a#123, 6#163, __auto_generated_subquery_name.a#164, __auto_generated_subquery_name.c#165, gid#162, __auto_generated_subquery_name.b#166]
      +- InMemoryRelation [a#123, b#124, c#125], StorageLevel(disk, memory, deserialized, 1 replicas)
            +- LocalTableScan [a#6, b#7, c#8]

This is because the literal 6 was used in the group-by expressions (referred to as gb in the query, and renamed 6#163 in the Expand operator's output attributes).

After this PR, foldable expressions in the aggregate expressions are kept as-is.

Why are the changes needed?

Some expressions require a foldable argument. In the above example, the round function requires a foldable expression as the scale argument. Because the scale argument is patched with an attribute, RoundBase#checkInputDataTypes returns an error, which leaves the Aggregate operator unresolved:

[INTERNAL_ERROR] Invalid call to dataType on unresolved object SQLSTATE: XX000
org.apache.spark.sql.catalyst.analysis.UnresolvedException: [INTERNAL_ERROR] Invalid call to dataType on unresolved object SQLSTATE: XX000
	at org.apache.spark.sql.catalyst.analysis.UnresolvedAttribute.dataType(unresolved.scala:255)
	at org.apache.spark.sql.catalyst.types.DataTypeUtils$.$anonfun$fromAttributes$1(DataTypeUtils.scala:241)
	at scala.collection.immutable.List.map(List.scala:247)
	at scala.collection.immutable.List.map(List.scala:79)
	at org.apache.spark.sql.catalyst.types.DataTypeUtils$.fromAttributes(DataTypeUtils.scala:241)
	at org.apache.spark.sql.catalyst.plans.QueryPlan.schema$lzycompute(QueryPlan.scala:428)
	at org.apache.spark.sql.catalyst.plans.QueryPlan.schema(QueryPlan.scala:428)
	at org.apache.spark.sql.execution.SparkPlan.executeCollectPublic(SparkPlan.scala:474)
        ...

Does this PR introduce any user-facing change?

No.

How was this patch tested?

New tests.

Was this patch authored or co-authored using generative AI tooling?

No.

[bersprockets]

Note this does result in a slight (but AFAIK harmless) oddity in the plan when the grouping expressions are included in the select statement, e.g.:

create or replace temp view v1(a, b, c) as values
(1, 1.001d, 2), (2, 3.001d, 4), (2, 3.001, 4);

cache table v1;

select
  a,
  gb,
  round(sum(b), 6) as sum1,
  count(distinct a) as count1,
  count(distinct c) as count2
from (
  select 6 as gb, *
  from v1
)
group by a, gb;

The aggregate operator will look like this:

Aggregate [a#257, 6#294], [a#257, 6 AS gb#244, round(first(sum(__auto_generated_subquery_name.b)#298, true) FILTER (WHERE (gid#293 = 0)), 6) AS sum1#245, count(__auto_generated_subquery_name.a#295) FILTER (WHERE (gid#293 = 1)) AS count1#246L, count(__auto_generated_subquery_name.c#296) FILTER (WHERE (gid#293 = 2)) AS count2#247L]

In the list of grouping expressions, gb will show up as the attribute 6#294 (taken from the Expand operator's output), but gb will show up in the aggregate expressions as 6 AS gb#244.

[dongjoon-hyun]

nit. Could you capitalize SQL syntax, @bersprockets ?

[dongjoon-hyun]

+1, LGTM.

cc @cloud-fan , @yaooqinn , @itholic (as a release manager of Apache Spark 3.5.3)

[bersprockets]

By the way, this is not a recent regression. I can reproduce on 3.2.2. It's possible it's always been this way.

[dongjoon-hyun]

Merged to master/3.5/3.4.

[cloud-fan]

late LGTM

[yaooqinn]

late LGTM