In coming up with a reference example UDF against 1.0.9 code, i created a json pointer example. For the mariadb server side implementation you must declare the return type so i made this string and you have one method implementation. The distributed columnstore udf requires you to implement a variation per type. In my case clearly the string type was needed and mostly used. I tried varying permutations and i found that in this case i only really needed to implement the getDouble and getBoolean variants when using the function in a where clause where the right side was a constant or other numeric / boolean column. It's unclear where the decision point is on determining the type but always choosing double seems odd. We should clearly map out for each case what the conversion logic is and possibly address any case we determine as a bug and document this better.