`GROUP BY CUBE/ROLLUP` over an ARRAY column with an array-lambda predicate in HAVING panics with `UnwindError 1104` (BadDataValueType downcast failure) — both members of a logically-equivalent query pair fail

[Manuel-Neuer1]

Search before asking

• I had searched in the issues and found no similar issues.

Version

8.0.90-v1.2.925-patch-10-1abb4b1e29 (rust nightly 2026-08-25) — the latest patch release as of 2026-08-25. Reproduced with the minimized form 4/4 runs (the original, more complex predicate shape reproduced nondeterministically).

What's Wrong?

When an ARRAY column is used as a GROUP BY key of a super-aggregate grouping (CUBE / ROLLUP),
and the HAVING predicate evaluates an array lambda (ARRAY_FILTER) over that same column, the query
fails with an internal downcast panic instead of evaluating the lambda with NULL propagation:

ERROR 1105 (HY000): UnwindError. Code: 1104, Text = called `Result::unwrap()` on an `Err` value:
BadDataValueType. Code: 1010, Text = failed to downcast column
NullableColumn { column: ArrayColumn { values: NullableColumn { column: Int32([19]),
validity: [0b_______1] }, offsets: [0, 1] }, validity: [0b_______1] }
into databend_common_expression::types::array::ArrayType<...::generic::GenericType<0>>.

The failing value is the NULL-filled group key of a super-aggregate (sub-total) row — CUBE/ROLLUP
fill the rolled-up grouping levels with NULL, and the array-lambda evaluation attempts to downcast the
NULL-containing ArrayColumn into a concrete ArrayType and unwraps the Err.

How to Reproduce?

CREATE TABLE t1(c0array ARRAY(INT) NOT NULL);
INSERT INTO t1(c0array) VALUES ([17, 27]), ([46, 14, 34]), ([-23, 37, 9, -20]), ([19]);

The two queries below are logically equivalent: the second is the standard predicate-relocation
rewrite of the first — the HAVING predicate is materialized as a projection column of a derived table
and filtered by an outer WHERE (π(σ(π(R))) ≡ π(σ(R)), the same equivalence every
predicate-pushdown optimizer relies on). A correct engine must return the same rows from both forms.

-- Base form (predicate in HAVING): fails
SELECT COALESCE(CAST(CAST(t1.c0array AS VARIANT) AS VARCHAR), '__NULL__') FROM t1
GROUP BY CUBE (t1.c0array)
HAVING (ARRAY_LENGTH(ARRAY_FILTER(t1.c0array, x -> x <> -42)) >= 0);
-- ERROR 1105 (HY000): UnwindError. Code: 1104, ... failed to downcast column ...

-- Relocated form (same predicate materialized into the derived-table projection, filtered outside):
-- fails with the SAME error
SELECT ref0 FROM (
    SELECT COALESCE(CAST(CAST(t1.c0array AS VARIANT) AS VARCHAR) AS ref0,
           (ARRAY_LENGTH(ARRAY_FILTER(t1.c0array, x -> x <> -42)) >= 0) AS ref1
    FROM t1 GROUP BY CUBE (t1.c0array)
) AS s WHERE ref1;
-- ERROR 1105 (HY000): UnwindError. Code: 1104, ... failed to downcast column ...

Both members of the equivalent pair crash identically, which localizes the defect to the scalar
array-lambda evaluation over the super-aggregate NULL key — independent of where the predicate
sits (HAVING vs. materialized projection). That positional independence is itself the diagnostic
value of presenting the pair: a filter-path-only failure would instead point at predicate
pushdown/planning.

Isolation matrix (same table, same grouping, only the HAVING predicate varies):

HAVING predicate	Result
constant (1 = 1)	OK (5 groups)
GROUPING(c0array) = 0	OK (4 groups)
ARRAY_LENGTH(ARRAY_FILTER(c0array, x -> x <> -42)) >= 0	UnwindError, stable

Additional contrasts that pin the trigger on the super-aggregate NULL key:

-- Same lambda predicate, plain GROUP BY (no CUBE/ROLLUP): succeeds, returns the 4 groups
SELECT t1.c0array FROM t1
GROUP BY t1.c0array
HAVING (ARRAY_LENGTH(ARRAY_FILTER(t1.c0array, x -> x <> -42)) >= 0);

-- Same CUBE grouping without the lambda: succeeds, returns the 4 non-total groups
SELECT t1.c0array FROM t1
GROUP BY CUBE (t1.c0array)
HAVING (GROUPING(t1.c0array) = 0);

-- ROLLUP triggers it as well:
SELECT t1.c0array FROM t1
GROUP BY ROLLUP (t1.c0array)
HAVING (ARRAY_LENGTH(ARRAY_FILTER(t1.c0array, x -> x <> -42)) >= 0);
-- ERROR 1105 (HY000): UnwindError. Code: 1104, ...

Are you willing to submit PR?

• Yes I am willing to submit a PR!