BUG: Parquet does not support saving fp16 GH#44846

doc/source/whatsnew/v1.4.0.rst

@@ -754,6 +754,7 @@ I/O
 - Bug in :func:`read_csv` when passing simultaneously a parser in ``date_parser`` and ``parse_dates=False``, the parsing was still called (:issue:`44366`)
 - Bug in :func:`read_csv` silently ignoring errors when failling to create a memory-mapped file (:issue:`44766`)
 - Bug in :func:`read_csv` when passing a ``tempfile.SpooledTemporaryFile`` opened in binary mode (:issue:`44748`)
+- Bug in :func:`to_parquet` where ``engine="pyarrow"`` could result in partial write when column dtype is ``float16``. (:issue:`44846)
 -
 
 Period

pandas/io/parquet.py

@@ -167,6 +167,13 @@ def write(
     ):
         self.validate_dataframe(df)
 
+        fp16_columns = df.select_dtypes(include="float16").columns
+        if fp16_columns.size > 0:
+            raise ValueError(
+                f"Columns [{','.join(fp16_columns.values)}] are of dtype float16. "
+                + "PyArrow does not support saving float16 dtype columns."
+            )
+
         from_pandas_kwargs: dict[str, Any] = {"schema": kwargs.pop("schema", None)}
         if index is not None:
             from_pandas_kwargs["preserve_index"] = index

pandas/tests/io/test_parquet.py

@@ -715,6 +715,18 @@ def test_duplicate_columns(self, pa):
         df = pd.DataFrame(np.arange(12).reshape(4, 3), columns=list("aaa")).copy()
         self.check_error_on_write(df, pa, ValueError, "Duplicate column names found")
 
+    def test_unsupported_float16(self, pa):
+        # #44847
+        # Not able to write float 16 column using pyarrow.
+        data = np.arange(2, 10, dtype=np.float16)
+        df = pd.DataFrame(data=data, columns=["fp16"])
+        fp16_columns = fp16_columns = df.select_dtypes(include="float16").columns
+        msg = (
+            f"Columns \\[{','.join(fp16_columns.values)}\\] are of dtype float16. "
+            + "PyArrow does not support saving float16 dtype columns."
+        )
+        self.check_error_on_write(df, pa, ValueError, msg)
+
     def test_unsupported(self, pa):
         # timedelta
         df = pd.DataFrame({"a": pd.timedelta_range("1 day", periods=3)})