[FEA] Allow input split on CUDF too long exceptions

[revans2]

Is your feature request related to a problem? Please describe.
Once #7866 goes in we will have the ability to split input data when executing expressions. But one of the limitations with CUDF is that of a int as the index in offsets. This limits the maximum size any column can be, including strings and lists/arrays. This is especially problematic for deeply nested types.

We should have a way to catch an exception from cudf and split/retry the operation if that exception is one that indicates that the output was too long for CUDF to support. We are likely going to have to make changes to CUDF and to the CUDF JNI APIs so it is easy to tell when this happens.

[ttnghia]

I think we can implement a specific exception class in the Java JNI layer. Then we have a dedicated API in JNI to check for column size. Something like:

static void throwIfSizeExceedCudfLimit(...)
  if (inputColumn.numRows() > long(INT_MAX)) {
    throw new ColumnSizeExceedCudfLimitException(...);
  }
}

The plugin just try+catch that specific exception for split+retry.

Probably that throwIfSizeExceedCudfLimit check is called in ColumnVector (ColumnView?) constructor so we always have new column being checked.

[revans2]

The problem is not with the input, but the output of operations. CUDF is the one that hopefully finds that the output is too large and throws an exception.

https://github.com/rapidsai/cudf/blob/2969b241c0654a11d1a61e29664bcaecd7bc4a15/cpp/include/cudf/strings/detail/strings_children.cuh#L82-L84

But that is not guaranteed in all cases because of overflow. The ideal would be that when a string, or any other column type, would exceed the CUDF limits then there could be an exception class that is specific to this so we could know that it happened and then retry the operation with a smaller input to avoid the issue.

[ttnghia]

I see. Recently cudf introduced several new exception class (rapidsai/cudf#12426) to throw in case of invalid input type. It should be reasonable to add a new exception type for our specific need.

I've filed a related issue: rapidsai/cudf#12925.