NaN handling

Float columns

Float NaN values are handled via IEEE 754 semantics (x != x is true iff NaN). Note the asymmetry with the column-form helpers below: float helpers operate on tensors (extract the column with get_float_col first), while the _col helpers take a Frame plus a column name.

• is_nan(col): returns a bool tensor (true = missing)
• fill_nan(col, fill_val): replace NaN with fill_val in a float tensor
• drop_nan(df, col_name): remove rows where the column is NaN
• any_nan(col): true if any value is NaN
• count_nan(col): count of NaN values

Integer columns

Integer columns carry an explicit boolean missing-value mask (true = missing). The mask is propagated through joins (sentinel rows), aggregations (masked rows are skipped), and concat. Use the _col variants:

• is_nan_col(df, col_name): returns the bool mask tensor
• fill_nan_col(df, col_name, fill_val): replace masked entries with fill_val
• drop_nan_col(df, col_name): remove rows where the mask is true
• any_nan_col(df, col_name): true if any entry is masked
• count_nan_col(df, col_name): count of masked entries

To construct an int column directly: int_col_of_list([1, 2, 3]) creates an IntCol with an all-false (no missing) mask.