[FEA] JNI support get map value for all value types and for all basic key types #9109
Comments
revans2
added
feature request
sameerz
changed the title
beckernick
added
libcudf
|
This issue has been labeled |
|
We are exploring doing this a different way, but for now we still want it. |
|
This issue has been labeled |
|
Same comment as before |
|
I've bumped this to 22.04 for now. |
…0380) Fixes #9109. This commit adds a `map` abstraction over a `column_view` of type `LIST<STRUCT<K,V>>`, where `K` and `V` are key and value types. A list column of structs with two members may thus be viewed as a `map` column. `maps_column_view` is to a `LIST<STRUCT<K,V>>` column what `lists_column_view` is to a `LIST` column. The `maps_column_view` abstraction provides methods to fetch lists of keys and values (as `LIST<K>` and `LIST<V>` respectively). It also provides map lookup methods to find the values corresponding to a specified key, for each row in the "map" column. E.g. ```c++ auto input_column = get_list_of_structs_col(); // input_column == [ {1:10, 2:20}, {1:100, 3:300}, {2:2000, 3:3000, 4:4000} ]; auto maps_view = cudf::jni::maps_column_view{input_column->view()}; auto keys = maps_view.keys(); // keys == [ {1,2}, {1,3}, {2,3,4} ]; auto values = maps_view.values(); // values == [ {10,20}, {100, 300}, {2000, 3000, 4000} ]; auto lookup_1 = maps_view.get_values_for( *make_numeric_scalar(1) ); // lookup_1 = [ {10, 100, null} ]; ``` This abstraction should help replace the Java/JNI `map_lookup` and `map_contains` kernels, which only handles `MAP<STRING, STRING>`. Authors: - MithunR (https://github.com/mythrocks) Approvers: - Jason Lowe (https://github.com/jlowe) - AJ Schmidt (https://github.com/ajschmidt8) - Nghia Truong (https://github.com/ttnghia) - Jake Hemstad (https://github.com/jrhemstad) URL: #10380
Is your feature request related to a problem? Please describe.
We would like to be able to support getting map values on more than just string to string maps. CUDF does not officially support maps and in JNI we have a stand alone special API just for string to string get map values. This is to update that existing code unless we can add some list API operations to support it more generally.
Describe the solution you'd like
Update the existing JNI cudf code for get map values to be more generic. Longer term support more list operations so we can do this correctly.
Describe alternatives you've considered
It would be great to have a list_position. Which would return the position of the first element in a list that is equal to a value provided as either a scalar or a column vector.
We could then create map get from this by pulling out the keys from the map and calling list_position on it, and then pulling out the values list column and calling extract_list_element on it and the result of list_position.
The text was updated successfully, but these errors were encountered: