Output size assertions in unify_json_strings

[jlowe]

unify_json_strings in map_utils.cu has two places where it asserts on string column data sizes. The first is this one:

  auto const output_size =
    2l +                                            // two extra bracket characters '[' and ']'
    static_cast<int64_t>(chars_size) +
    static_cast<int64_t>(input.size() - 1) +        // append `,` character between input rows
    static_cast<int64_t>(input.null_count()) * 2l;  // replace null with "{}"
  CUDF_EXPECTS(output_size <= static_cast<int64_t>(std::numeric_limits<cudf::size_type>::max()),
               "The input json column is too large and causes overflow.");

Eventually cudf will support string columns with more than 2GB of character data, and this assert will fire when we do not want it to.

Later there's a sanity check assertion that used to be super cheap, because before the recent cudf strings column change, calculating the size of character data in a string column involved only accessing data already on the CPU. Now this assert requires an extra stream synchronization that could be avoided by simply computing it directly from the known output_size.

  auto const joined_input_size_bytes = joined_input_scv.chars_size(stream);
  CUDF_EXPECTS(joined_input_size_bytes + 2 == output_size, "Incorrect output size computation.");

We could consider removing this assertion if we never see issues with join_strings doing what it's supposed to do, and that would remove a stream synchronization required to access the string column's character data size.

[sameerz]

Keeping this in the backlog until libcudf has long strings support per rapidsai/cudf#13733