[FEA] Support left-semi and left-anti joins in `cudf::hash_join`

[gaohao95]

Is your feature request related to a problem? Please describe.

cudf::hash_join makes it possible to build the hash table once and probe it multiple times. But it only supports inner join, left join and full join. I wish cudf::hash_join can support left-semi and left-anti join as well.

[GregoryKimball]

Thank you @gaohao95 for suggesting this. We will do some scoping and return to this request.

[vyasr]

There are some important limitations to be aware of.

• Conceptually, semi/anti joins actually use a different data structure than inner/left/full joins (maps vs multimaps). The reason is that with inner/left/full joins each row in one table could corresponding to multiple rows in the other table, i.e. the multimap supports non-unique keys. Therefore, the reuse would be limited between two disjoint sets of APIs: semi_join could reuse a map built for anti_join and vice versa, but it could not use a multimap built for inner/left/full joins.
• The current implementation of semi/anti joins relies internally on cudf::contains, which is what builds the hash map internally. Therefore, beyond the previous conceptual limitation, supporting this functionality would also require a deeper refactor to expose the hash table in some way.
• There is ongoing work to refactor cuco data structures and expand their usage within libcudf. I would not recommend making any changes to the join APIs until that work is further along.

[gaohao95]

Thanks @vyasr! Those are good points!

Therefore, the reuse would be limited between two disjoint sets of APIs: semi_join could reuse a map built for anti_join and vice versa, but it could not use a multimap built for inner/left/full joins.

In my use case (broadcast join) this is fine. An object is only needed to probe a single join type.

There is ongoing work to refactor cuco data structures and expand their usage within libcudf. I would not recommend making any changes to the join APIs until that work is further along.

This is not a blocker for us so we can wait.

[ahmet-uyar]

We also needed this recently for a broadcast join implementation. We would prefer if we cudf::hash_join supports left-semi join.