fix(optimizer)!: stop pruning implicit GROUP BY ALL keys during projection pushdown

[fivetran-kwoodbeck]

pushdown_projections prunes columns that aren't referenced by an outer query. But a bare GROUP BY ALL infers its grouping keys from projections in the SELECT list, so pruning an "unused" column silently changes which columns are grouped by, corrupting the aggregation.

Input:

SELECT t.a FROM (SELECT a, b, SUM(b) AS s FROM x GROUP BY ALL) t

Wrong output (b and s are dropped):

SELECT "t"."a" AS "a" FROM (SELECT "x"."a" AS "a" FROM "x" AS "x" GROUP BY ALL) AS "t"

Expected output

SELECT "t"."a" AS "a" FROM (SELECT "x"."a" AS "a", "x"."b" AS "b" FROM "x" AS "x" GROUP BY ALL) AS "t"

Fix

We now check whether the scope has a GROUP BY ALL and keeps projections regardless of outer references.

[github-actions]

SQLGlot Integration Test Results

❌ 1 regression — see details below

Comparing:

• this branch (sqlglot:optimizer/fix-pushdown-prunes-group-by-all-keys @ sqlglot 9d9796c)
• baseline (main @ sqlglot 194e12b)

By Dialect

dialect	main	feature branch	transitions	links
bigquery -> bigquery	34287/34695 passed (98.8%)	33136/33540 passed (98.8%)	1 pass -> fail	full result / delta

Overall

main: 182937 total, 160863 passed (pass rate: 87.9%)

sqlglot:optimizer/fix-pushdown-prunes-group-by-all-keys: 170743 total, 149710 passed (pass rate: 87.7%)

Transitions:
1 pass -> fail

Dialect pair changes: 0 previous results not found, 3 current results not found

❌ 1 regression (view logs)

[georgesittas]

LGTM, one nit re: reintroducing short-circuiting on the AggFunc search.

@fivetran-kwoodbeck the same bug exists for ORDER BY ALL:

-- input
SELECT t.a, t.c FROM (SELECT a, b, c FROM x ORDER BY ALL LIMIT 1) t
-- optimizer output: b is pruned
WITH "t" AS (SELECT "x"."a" AS "a", "x"."c" AS "c" FROM "x" AS "x" ORDER BY ALL LIMIT 1) ...

• This is only observable when the subquery's ordering matters, i.e. with LIMIT/OFFSET
• Unlike GROUP BY ALL, there's no aggregate exemption: with ORDER BY ALL every projection is a sort key
• We should check for a Var presence wrapping "ALL" to capture dialects that actually support this

[georgesittas]

I'll get this in and do a quick clean up - I wanna release shortly.