fix: unnest_subqueries produces invalid column references when subquery contains a UNION

[snovik75]

Fixes #7666

Problem

When unnest_subqueries encounters a SetOperation (UNION/UNION ALL) subquery, it wraps it in a derived table:

if isinstance(select, exp.SetOperation):
    select = exp.select(*select.selects).from_(select.subquery(next_alias_name()))

select.selects returns the fully-qualified column expressions from the left branch of the UNION (e.g. child_table.col_a AS col_a). These are copied verbatim into the outer wrapper SELECT, but the FROM source is the new subquery alias _u_0, making the column references invalid in that scope.

Fix

Build column references that point to the derived alias instead of copying the inner expressions:

if isinstance(select, exp.SetOperation):
    inner_alias = next_alias_name()
    select = (
        exp.select(*(
            exp.alias_(exp.column(s.alias_or_name, inner_alias), s.alias_or_name)
            for s in select.selects
        ))
        .from_(select.subquery(inner_alias))
    )

Reproducer

import sqlglot
from sqlglot.optimizer.qualify import qualify
from sqlglot.optimizer.unnest_subqueries import unnest_subqueries

sql = """
SELECT t.id AS ref_id
FROM parent_table t
WHERE t.id NOT IN (
    SELECT DISTINCT col_a FROM child_table
    UNION ALL
    SELECT col_b FROM child_table
)
"""

ast = sqlglot.parse_one(sql)
result = unnest_subqueries(qualify(ast))
print(result.sql(pretty=True))

Before: child_table.col_a referenced inside a scope where only _u_0 exists.

After: _u_0.col_a — correctly qualified against the derived alias.

[snovik75]

dear team, there was a but in the old test. could you please have a careful look? thanks

[snovik75]

tbh, that test in optimizer.sql seems redundant and/or misplaced (and does not have a title) but I just corrected it. your call what to do with it

[geooo109]

LGTM

A note here (future work):

CREATE TABLE x AS SELECT * FROM (VALUES (1),(2),(3),(4)) AS t(a);
CREATE TABLE y AS SELECT * FROM (VALUES (1),(NULL)) AS t(a);
CREATE TABLE z AS SELECT * FROM (VALUES (2),(3)) AS t(a);
SELECT * FROM x WHERE x.a NOT IN (SELECT y.a AS a FROM y UNION ALL SELECT z.a AS a FROM z);

For this case ^ in duckdb the optimized query doesn't match the semantics of the input.

[georgesittas]

Thanks!

[georgesittas]

This is embarassing, seems like I didn't notice this at all when I added the test...