[VL] Incorrect result with small Parquet pages + Velox aggregation pushdown

[zhztheplayer]

Backend

VL (Velox)

Bug description

Found in Gluten + Velox when running ClickBench Q31. This is reproduce-able with the default Gluten / Spark configurations. No specific options are required to set.

The minimal SQL to reproduce:

SELECT SearchEngineID, SUM(IsRefresh) FROM hits WHERE SearchPhrase <> '' AND (SearchEngineID in (1, 2, 4)) AND ClientIP = -807147100 GROUP BY SearchEngineID

Data:

hits.parquet from https://github.com/ClickHouse/ClickBench/?tab=readme-ov-file#data-loading

Spark answer:

+--------------+--------------+
|SearchEngineID|sum(IsRefresh)|
+--------------+--------------+
|             1|             1|
|             2|            35|
|             4|             3|
+--------------+--------------+

Gluten (Velox) answer:

+--------------+--------------+
|SearchEngineID|sum(IsRefresh)|
+--------------+--------------+
|             4|             1|
|             2|            37|
|             1|             1|
+--------------+--------------+

Gluten version

Latest

Spark version

3.5

[zhztheplayer]

This is a Velox bug. Fix facebookincubator/velox#14332.