SparkUtils.flattenSchema method throws null pointer exception

Describe the bug

SparkUtils.flattenSchema method throws NPE for the empty data frame having occurs or schema contains arrays

To Reproduce

create a data frame from empty data file and copybook with occurs and flatten the schema with SparkUtils.flattenSchema method

Expected behaviour

It should return the flatten the schema with zero records as it's empty dataset

Screenshots

Additional context

LOC which is causing NPE

var maxInd = df.agg(max(expr(s"size($path${structField.name})"))).collect()(0)(0).toString.toInt

in function flattenStructArray

I understand the above code depends on data because of OCCURS DEPENDING ON implementation and it will throw NPE in case for zero records in dataframe

Potential Fix

Approach 1: Before calculating maxInd We can check if dataframe is empty. If its empty we can get the max index of a group from copybook for OCCURS or OCCURS DEPENDING on.For empty dataset it is not required to get the maxInd from the data

This will return the dataframe with maximum possible columns from Occurs or OCCOURS DEPENDING on.
If it is OCCURS DEPENDING ON FROM 0 to 3 then we will consider 3. We have to pass copybookcontents in this case to the required function.

Approach 2:
We can only update the code to read the maxInd from copybook in case of OCCURS. For OCCURS DEPENDING ON we can use the existing code.

As per my view Approach 1 is good. Kindly let me know your view.

[yruslan]

Thanks for the report! We will fix the NPE exception.

The issue with getting info from OCCURS is once the Cobol schema is converted to Spark schema the maximum array size is lost because Spark arrays do not have maximum number of elements as part of metadata. That's why the maximum array elements are determined by the actual data.

I have some ideas, will let you know after I try them.

Sure thanks. Please check once Approach 1 where we can pass optional parameter as copybook contents and can derive max occurrences based on that.

for the below copybook in case of zero byte file the dataset will be generated as mentioned below.

  01 RECORD.
      02 COUNT PIC 9(1).
      02 GROUP OCCURS 0 TO 2 TIMES DEPENDING ON COUNT.
         03 INNER-COUNT PIC 9(1).
         03 INNER-GROUP OCCURS 0 TO 3 TIMES
                            DEPENDING ON INNER-COUNT.
            04 FIELD PIC X.

-----+-------------------+---------------------------+---------------------------+---------------------------+-------------------+---------------------------+---------------------------+---------------------------+
|COUNT|GROUP_0_INNER_COUNT|GROUP_0_INNER_GROUP_0_FIELD|GROUP_0_INNER_GROUP_1_FIELD|GROUP_0_INNER_GROUP_2_FIELD|GROUP_1_INNER_COUNT|GROUP_1_INNER_GROUP_0_FIELD|GROUP_1_INNER_GROUP_1_FIELD|GROUP_1_INNER_GROUP_2_FIELD|
+-----+-------------------+---------------------------+---------------------------+---------------------------+-------------------+---------------------------+---------------------------+---------------------------+
+-----+-------------------+---------------------------+---------------------------+---------------------------+-------------------+---------------------------+---------------------------+---------------------------+

[yruslan]

The idea worked. When creaing a Spark schema from a copybook 'minElements' and 'maxElements' metadata fields are added to arrays from OCCURS. This way the program no longer needs to determine the maximum array size from the data itself. If metadata fields are not available the program will get these maximums the old ways - through querying the data.

You can check the update in master, or wait until 2.4.8 s released.

@yruslan Thanks a lot. Any tentative date for 2.4.8 release.

[yruslan]

Probably until the end of the week.

But I would encourage you to check if the updated flattenSchema() works for you from the current master, so if not changes could be made before the release.

sure let me pull the changes and check via UTCs

As we are currently on 2.1.3 which uses spark 2.4.5 I suppose the newer version should be backward compatible as currently we may not use spark 3.x