fix: Fix out nullability for structs coming from arrow

crates/polars-core/src/chunked_array/struct_/mod.rs

@@ -301,7 +301,7 @@ impl StructChunked {
     }
 
     /// Set the outer nulls into the inner arrays, and clear the outer validity.
-    fn propagate_nulls(&mut self) {
+    pub(crate) fn propagate_nulls(&mut self) {
         if self.null_count > 0 {
             // SAFETY:
             // We keep length and dtypes the same.

crates/polars-core/src/series/from.rs

@@ -103,8 +103,10 @@ impl Series {
             BinaryOffset => BinaryOffsetChunked::from_chunks(name, chunks).into_series(),
             #[cfg(feature = "dtype-struct")]
             Struct(_) => {
-                StructChunked::from_chunks_and_dtype_unchecked(name, chunks, dtype.clone())
-                    .into_series()
+                let mut ca =
+                    StructChunked::from_chunks_and_dtype_unchecked(name, chunks, dtype.clone());
+                ca.propagate_nulls();
+                ca.into_series()
             },
             #[cfg(feature = "object")]
             Object(_, _) => {
@@ -400,11 +402,12 @@ impl Series {
             #[cfg(feature = "dtype-struct")]
             ArrowDataType::Struct(_) => {
                 let (chunks, dtype) = to_physical_and_dtype(chunks, md);
+
                 unsafe {
-                    Ok(
-                        StructChunked::from_chunks_and_dtype_unchecked(name, chunks, dtype)
-                            .into_series(),
-                    )
+                    let mut ca =
+                        StructChunked::from_chunks_and_dtype_unchecked(name, chunks, dtype);
+                    ca.propagate_nulls();
+                    Ok(ca.into_series())
                 }
             },
             ArrowDataType::FixedSizeBinary(_) => {

py-polars/tests/unit/datatypes/test_struct.py

@@ -959,3 +959,9 @@ def test_struct_chunked_gather_17603() -> None:
             {"a": 3},
         ]
     }
+
+
+def test_struct_out_nullability_from_arrow() -> None:
+    df = pl.DataFrame(pd.DataFrame({"abc": [{"a": 1.0, "b": pd.NA}, pd.NA]}))
+    res = df.select(a=pl.col("abc").struct.field("a"))
+    assert res.to_dicts() == [{"a": 1.0}, {"a": None}]