fix(rust, python): correct invalid type in struct anyvalue access (#5844)

Author: ritchie46

polars/polars-core/src/chunked_array/ops/any_value.rs

@@ -112,10 +112,16 @@ impl<'a> AnyValue<'a> {
             AnyValue::Struct(idx, arr, flds) => {
                 let idx = *idx;
                 unsafe {
-                    arr.values()
-                        .iter()
-                        .zip(*flds)
-                        .map(move |(arr, fld)| arr_to_any_value(&**arr, idx, fld.data_type()))
+                    arr.values().iter().zip(*flds).map(move |(arr, fld)| {
+                        // TODO! this is hacky. Investigate if we only should put physical types
+                        // into structs
+                        if let Some(arr) = arr.as_any().downcast_ref::<DictionaryArray<u32>>() {
+                            let keys = arr.keys();
+                            arr_to_any_value(keys, idx, fld.data_type())
+                        } else {
+                            arr_to_any_value(&**arr, idx, fld.data_type())
+                        }
+                    })
                 }
             }
             _ => unreachable!(),

polars/polars-ops/src/series/ops/various.rs

@@ -13,7 +13,7 @@ pub trait SeriesMethods: SeriesSealed {
         let groups = s.group_tuples(multithreaded, sorted)?;
         let values = unsafe { s.agg_first(&groups) };
         let counts = groups.group_lengths("counts");
-        let cols = vec![values.into_series(), counts.into_series()];
+        let cols = vec![values, counts.into_series()];
         let df = DataFrame::new_no_checks(cols);
         if sorted {
             df.sort(["counts"], true)

py-polars/tests/unit/test_struct.py

@@ -704,3 +704,17 @@ def test_struct_any_value_get_after_append() -> None:
     a = a.append(b)
     assert a[0] == {"a": 1, "b": 2}
     assert a[1] == {"a": 2, "b": 3}
+
+
+def test_struct_categorical_5843() -> None:
+    df = pl.DataFrame({"foo": ["a", "b", "c", "a"]}).with_column(
+        pl.col("foo").cast(pl.Categorical)
+    )
+    result = df.select(pl.col("foo").value_counts(sort=True))
+    assert result.to_dict(False) == {
+        "foo": [
+            {"foo": "a", "counts": 2},
+            {"foo": "b", "counts": 1},
+            {"foo": "c", "counts": 1},
+        ]
+    }