[VL] Native Scan operator incurred inaccuracy in some filter cases. #8764
Description
Backend
VL (Velox)
Bug description
CREATE TABLE default.testtable (
a bigint,
b bigint,
c int
)
USING parquet
LOCATION 'file:///data/dataset/testtable';
dataset:
query:
select
max(a) as id
from
testtable
where
c = 2
group by
b;
[Expected behavior]
spark-sql (default)> select
> max(a) as id
> from
> testtable
> where
> c = 2
> group by
> b;
299
275
233
237
323
321
253
102
226
219
297
319
327
249
96
Time taken: 4.221 seconds, Fetched 15 row(s)
[Actual behavior]
spark-sql (default)> select
> max(a) as id
> from
> testtable
> where
> c = 2
> group by
> b;
139938755249256
139938755249256
139936948064087
139945378802512
139945378802512
139938755303160
139938755303160
139945378802512
139945378802512
139938755249256
139945378802512
139945378802512
139945378802512
139938755249256
139938755303160
Time taken: 4.244 seconds, Fetched 15 row(s)
This error can be reproduced consistently.
But if Velox is compiled in DEBUG mode, the result is also accurate.
cc @weiting-chen @zhouyuan , Thanks.