fix: Reading of reencoded categorical in Parquet (#22436)

Author: coastalwhite

crates/polars-parquet/src/arrow/read/deserialize/simple.rs

@@ -387,9 +387,13 @@ pub fn page_iter_to_array(
             // - Int -> String which can be turned into categoricals
             assert_eq!(value_type.as_ref(), &ArrowDataType::Utf8View);
 
-            if field.metadata.is_none_or(|md| {
-                !md.contains_key(DTYPE_ENUM_VALUES) && !md.contains_key(DTYPE_CATEGORICAL)
-            }) {
+            if field.metadata.is_some_and(|md| {
+                md.contains_key(DTYPE_ENUM_VALUES) || md.contains_key(DTYPE_CATEGORICAL)
+            }) && matches!(key_type, IntegerType::UInt32)
+            {
+                PageDecoder::new(pages, dtype, CategoricalDecoder::new(), init_nested)?
+                    .collect_boxed(filter)?
+            } else {
                 let (nested, array, ptm) = PageDecoder::new(
                     pages,
                     ArrowDataType::Utf8View,
@@ -404,10 +408,6 @@ pub fn page_iter_to_array(
                         .unwrap(),
                     ptm,
                 )
-            } else {
-                assert_eq!(key_type, &IntegerType::UInt32);
-                PageDecoder::new(pages, dtype, CategoricalDecoder::new(), init_nested)?
-                    .collect_boxed(filter)?
             }
         },
         (from, to) => {

crates/polars-parquet/src/arrow/read/schema/metadata.rs

@@ -1,5 +1,5 @@
 use arrow::datatypes::{
-    ArrowDataType, ArrowSchema, DTYPE_CATEGORICAL, DTYPE_ENUM_VALUES, Field, Metadata,
+    ArrowDataType, ArrowSchema, DTYPE_CATEGORICAL, DTYPE_ENUM_VALUES, Field, IntegerType, Metadata,
 };
 use arrow::io::ipc::read::deserialize_schema;
 use base64::Engine as _;
@@ -27,7 +27,8 @@ fn convert_field(field: &mut Field) {
         ArrowDataType::Dictionary(key_type, value_type, sorted) => {
             let is_pl_enum_or_categorical = field.metadata.as_ref().is_some_and(|md| {
                 md.contains_key(DTYPE_ENUM_VALUES) || md.contains_key(DTYPE_CATEGORICAL)
-            });
+            }) && matches!(key_type, IntegerType::UInt32)
+                && matches!(value_type.as_ref(), ArrowDataType::Utf8View);
             let is_int_to_str = matches!(
                 value_type.as_ref(),
                 ArrowDataType::Utf8View | ArrowDataType::Utf8 | ArrowDataType::LargeUtf8

py-polars/tests/unit/io/test_parquet.py

@@ -3180,3 +3180,24 @@ def test_filter_nan_22289() -> None:
         lf.collect().filter(pl.col.a.is_nan()),
         lf.filter(pl.col.a.is_nan()).collect(),
     )
+
+
+def test_reencode_categoricals_22385() -> None:
+    tbl = pl.Series("a", ["abc"], pl.Categorical()).to_frame().to_arrow()
+    tbl = tbl.cast(
+        pa.schema(
+            [
+                pa.field(
+                    "a",
+                    pa.dictionary(pa.int32(), pa.large_string()),
+                    metadata=tbl.schema[0].metadata,
+                ),
+            ]
+        )
+    )
+
+    f = io.BytesIO()
+    pq.write_table(tbl, f)
+
+    f.seek(0)
+    pl.scan_parquet(f).collect()