[FEA] Report the number of rows read per file in libcudf's Parquet reader

[gaohao95]

Is your feature request related to a problem? Please describe. I wish libcudf's parquet reader reports the number of rows read per file.

Consider the following example,

  std::vector<std::string> file_paths;  // defined elsewhere
  std::vector<std::string> column_names;  // defined elsewhere

  auto source  = cudf::io::source_info(file_paths);
  auto options = cudf::io::parquet_reader_options::builder(source);
  options.columns(column_names);
  auto result = cudf::io::read_parquet(options);

Here, result is of type table_with_metadata, but the metadata doesn't contain the number of rows read from each file. I wish libcudf can add this functionality.

Describe the solution you'd like Report the number of rows read from each file in table_with_metadata.

Describe alternatives you've considered I have tried cudf::io::read_parquet_metadata out-of-band, like the following snippet.

  std::vector<cudf::size_type> rows_per_file;
  rows_per_file.reserve(file_paths.size());

  for (auto const& file_path : file_paths) {
    auto file_source = cudf::io::source_info(file_path);
    auto metadata    = cudf::io::read_parquet_metadata(file_source);
    rows_per_file.push_back(metadata.num_rows());
  }
  result.rows_per_file = std::move(rows_per_file);

But this has nontrivial overhead in my use case. I believe we can get it for free as part of the Parquet reading process, since the Parquet reader needs to decode the file footers anyway.