Jack Chen created SPARK-44448:
---------------------------------
Summary: Wrong results for dense_rank() <= k from
InferWindowGroupLimit and DenseRankLimitIterator
Key: SPARK-44448
URL: https://issues.apache.org/jira/browse/SPARK-44448
Project: Spark
Issue Type: Bug
Components: SQL
Affects Versions: 3.4.0
Reporter: Jack Chen
Top-k filters on a dense_rank() window function return wrong results, due to a
bug in optimization InferWindowGroupLimit, specifically in the code for
DenseRankLimitIterator.
Repro:
{code:java}
create or replace temp view t1 (p, o) as values (1, 1), (1, 1), (1, 2), (2, 1),
(2, 1), (2, 2);
select * from (select *, dense_rank() over (partition by p order by o) as rnk
from t1) where rnk = 1;{code}
Spark result:
{code:java}
[1,1,1]
[1,1,1]
[2,1,1]{code}
Correct result:
{code:java}
[1,1,1]
[1,1,1]
[2,1,1]
[2,1,1]{code}
The bug is in {{{}DenseRankLimitIterator{}}}, it fails to reset state properly
when transitioning from one window partition to the next. {{reset}} only resets
{{{}rank = 0{}}}, what it is missing is to reset {{{}currentRankRow = null{}}}.
This means that when processing the second and later window partitions, the
rank incorrectly gets incremented based on comparing the ordering of the last
row of the previous partition to the first row of the new partition.
This means that a dense_rank window func that has more than one window
partition and more than one row with dense_rank = 1 in the second or later
partitions can give wrong results when optimized.
({{{}RankLimitIterator{}}} narrowly avoids this bug by happenstance, the first
row in the new partition will try to increment rank, but increment it by the
value of count which is 0, so it happens to work by accident).
Unfortunately, tests for the optimization only had a single row per rank, so
did not catch the bug as the bug requires multiple rows per rank.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
