DataFusion drops grouped MIN/MAX rows with NULL sort keys under ORDER BY + LIMIT #22190

Describe the bug

Version

• Python: 3.12.3
• Python package: datafusion==53.0.0
• Arrow input via pyarrow==24.0.0
• Platform used for reproduction: Ubuntu 24.04 x86_64, Linux 6.17.0-23-generic

Minimal Reproducer

import pyarrow as pa
from datafusion import SessionContext

ctx = SessionContext()
batch = pa.RecordBatch.from_arrays(
    [
        pa.array(["a"], type=pa.string()),
        pa.array([None], type=pa.int64()),
    ],
    schema=pa.schema(
        [
            pa.field("g", pa.string(), nullable=True),
            pa.field("x", pa.int64(), nullable=True),
        ]
    ),
)
ctx.register_record_batches("t0", [[batch]])

base = "SELECT g, MIN(x) AS min_x FROM t0 GROUP BY g"
print(ctx.sql(f"SELECT min_x FROM ({base}) q LIMIT 20").to_pandas())
print(ctx.sql(f"SELECT min_x FROM ({base}) q ORDER BY min_x ASC NULLS LAST LIMIT 20").to_pandas())

Actual Result

Standalone reproducer output:

datafusion=53.0.0
pyarrow=24.0.0
control:
   min_x
0    NaN
top-k:
Empty DataFrame
Columns: [min_x]
Index: []
top-k record-batch rows=0
AssertionError: DataFusion dropped the group whose aggregate sort key is NULL

The control query returns the grouped row with a NULL aggregate result:

   min_x
0    NaN

The query with ORDER BY min_x ASC NULLS LAST LIMIT 20 returns no rows:

Empty DataFrame
Columns: [min_x]
Index: []

With two groups, one having MIN(x) = NULL and one having MIN(x) = 5, the ORDER BY ... LIMIT 20 query returns only the non-NULL sort-key group. The limit is larger than the number of groups, so no grouped row should be removed.

Expected Result

Both single-group queries should return one row. GROUP BY g forms a group for g = 'a'; MIN(x) is NULL because all values in that group are NULL. Ordering with NULLS LAST and a limit larger than the result cardinality should preserve the row.

Additional Context

I found this with a cross-backend differential test. In the generated case, pandas, Polars eager/lazy, DuckDB, SQLite, and an independent DSL reference all preserved the grouped row whose aggregate sort key is NULL. DataFusion was the only backend that returned an empty result.

I also ran targeted follow-up checks for the same pattern. The issue consistently reproduced for NULL MIN(x) ordered ascending and NULL MAX(x) ordered descending, while nearby controls such as plain NULL column top-k, NULL group key with non-NULL aggregate, and NULL SUM(x) / AVG(x) aggregate top-k preserved the row.

I can share the standalone reproducer, boundary-check script, and full generated artifacts if useful.

To Reproduce

No response

Expected behavior

No response

Additional context

No response