fix(python): skip empty row groups during stats gathering (#2172)

# Description
For some odd reason the pyarrow parquet writer will leave empty row
groups in the parquet file when it hits the max_row limit that's passed.
While grabbing the stats we were checking if all row_groups were having
stats added to them but these empty row groups had no stats so it causes
the whole file add action to get no stats recorded.

We now skip empty row groups while gathering the stats to prevent this.

In v0.15.2 we now also evaluate files with no stats mentioned as null
@roeap @rtyler not sure if this is entirely correct as well

# Related Issue(s)
- closes #2169

---------

Co-authored-by: Will Jones <willjones127@gmail.com>

python/deltalake/writer.py

@@ -666,7 +666,8 @@ def get_file_stats_from_metadata(
 
     def iter_groups(metadata: Any) -> Iterator[Any]:
         for i in range(metadata.num_row_groups):
-            yield metadata.row_group(i)
+            if metadata.row_group(i).num_rows > 0:
+                yield metadata.row_group(i)
 
     for column_idx in range(metadata.num_columns):
         name = metadata.row_group(0).column(column_idx).path_in_schema

python/tests/test_writer.py

@@ -1251,3 +1251,25 @@ def test_with_deltalake_schema(tmp_path: pathlib.Path, sample_data: pa.Table):
     )
     delta_table = DeltaTable(tmp_path)
     assert delta_table.schema().to_pyarrow() == sample_data.schema
+
+
+def test_write_stats_empty_rowgroups(tmp_path: pathlib.Path):
+    # https://github.com/delta-io/delta-rs/issues/2169
+    data = pa.table(
+        {
+            "data": pa.array(["B"] * 1024 * 33),
+        }
+    )
+    write_deltalake(
+        tmp_path,
+        data,
+        max_rows_per_file=1024 * 32,
+        max_rows_per_group=1024 * 16,
+        min_rows_per_group=8 * 1024,
+        mode="overwrite",
+    )
+    dt = DeltaTable(tmp_path)
+    assert (
+        dt.to_pyarrow_dataset().to_table(filter=(pc.field("data") == "B")).shape[0]
+        == 33792
+    )