You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Is your feature request related to a problem or challenge? Please describe what you are trying to do.
A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]
(This section helps Arrow developers understand the context and why for this feature, in addition to the what)
When we have an Anti Join(Left Anti or Right Anti), like A Left Anti Join B on (A.id = B.id), the real join conditions
should be A Left Anti Join B on ((Or(A.id = B.id), IsNull(A.id = B.id))), this can not be executed as HashJoins directly.
For left Anti Join, if the right side contains Null values, should return Empty Relation
For right Anti Join, if the left side contains Null values, should return Empty Relation.
For HashJoin, it is relatively easy to implement such logic when the partition mode is CollectLeft, there is challenge when the partition mode is Partitioned.
For the Anti Joins converted from EXCEPT clause, our implement is correct.
For the Anti Joins converted from NOT EXISTS, our implement should be correct, need to confirm.
For the Anti Joins converted from NOT IN, we have to handle Null cases as NAAJ.
It would not surprise me at all if there are bugs handling ANTI joins in DataFusion. As I recall there is a bunch of subtlety regarding nulls when there are multi-column equijoins as well and there are nulls in some but not all columns 🤔
For the Anti Joins converted from EXCEPT clause, our implement is correct. For the Anti Joins converted from NOT EXISTS, our implement should be correct, need to confirm. For the Anti Joins converted from NOT IN, we have to handle Null cases as NAAJ.
I find the Join in the datafusion with a attribute
/// If null_equals_null is true, null == null else null != null
pub null_equals_null: bool,
, can this resolve the Null-Aware in the HASH-JOIN?
For the Anti Joins converted from EXCEPT clause, our implement is correct. For the Anti Joins converted from NOT EXISTS, our implement should be correct, need to confirm. For the Anti Joins converted from NOT IN, we have to handle Null cases as NAAJ.
I find the Join in the datafusion with a attribute
/// If null_equals_null is true, null == null else null != null
pub null_equals_null: bool,
, can this resolve the Null-Aware in the HASH-JOIN?
Activity
mingmwang commentedon Nov 15, 2022
@alamb @jackwener @liukun4515
Please share your thoughts.
alamb commentedon Nov 15, 2022
I will respond tomorrow -- also cc @Dandandan who may be interested
mingmwang commentedon Nov 16, 2022
For the Anti Joins converted from EXCEPT clause, our implement is correct.
For the Anti Joins converted from NOT EXISTS, our implement should be correct, need to confirm.
For the Anti Joins converted from NOT IN, we have to handle Null cases as NAAJ.
alamb commentedon Nov 16, 2022
It would not surprise me at all if there are bugs handling ANTI joins in DataFusion. As I recall there is a bunch of subtlety regarding nulls when there are multi-column equijoins as well and there are nulls in some but not all columns 🤔
My recommendation to make improvements in this area is start by writing tests (perhaps borrow some / all of the duckdb tests): https://github.com/duckdb/duckdb/tree/master/test/sql/join to figure what, if anything, we are missing
It would be nice to have a proper data driven test runner -- I will write up a ticket and see if I can get anyone excited about this 🤔
alamb commentedon Nov 16, 2022
Wrote up my dream testing tool: #4248
liukun4515 commentedon Dec 9, 2022
I find the Join in the datafusion with a attribute
, can this resolve the Null-Aware in the HASH-JOIN?
mingmwang commentedon Dec 9, 2022
No.