Arrow2 now supports non-ordered projections in IPC. Fixes pola-rs#1761.

polars/polars-io/src/ipc.rs

@@ -119,11 +119,7 @@ impl<R: Read + Seek> IpcReader<R> {
     ) -> Result<DataFrame> {
         let rechunk = self.rechunk;
         let metadata = read::read_file_metadata(&mut self.reader)?;
-        let projection = projection.map(|x| {
-            let mut x = x.to_vec();
-            x.sort_unstable();
-            x
-        });
+        let projection = projection.map(|x| x.to_vec());
 
         let schema = if let Some(projection) = &projection {
             apply_projection(&metadata.schema, projection)
@@ -180,11 +176,7 @@ where
         let schema = &metadata.schema;
 
         if let Some(columns) = self.columns {
-            let mut prj = columns_to_projection(columns, schema)?;
-
-            // Ipc reader panics if the projection is not in increasing order, so sorting is the safer way.
-            prj.sort_unstable();
-            self.projection = Some(prj);
+            self.projection = Some(columns_to_projection(columns, schema)?);
         }
 
         let schema = if let Some(projection) = &self.projection {

py-polars/tests/io/test_ipc.py

@@ -39,24 +39,24 @@ def test_from_to_file(
 
 def test_select_columns() -> None:
     df = pl.DataFrame({"a": [1, 2, 3], "b": [True, False, True], "c": ["a", "b", "c"]})
-    expected = pl.DataFrame({"b": [True, False, True], "c": ["a", "b", "c"]})
+    expected = pl.DataFrame({"c": ["a", "b", "c"], "b": [True, False, True]})
 
     f = io.BytesIO()
     df.write_ipc(f)
     f.seek(0)
 
-    read_df = pl.read_ipc(f, columns=["b", "c"], use_pyarrow=False)
+    read_df = pl.read_ipc(f, columns=["c", "b"], use_pyarrow=False)
     assert expected.frame_equal(read_df)
 
 
 def test_select_projection() -> None:
     df = pl.DataFrame({"a": [1, 2, 3], "b": [True, False, True], "c": ["a", "b", "c"]})
-    expected = pl.DataFrame({"b": [True, False, True], "c": ["a", "b", "c"]})
+    expected = pl.DataFrame({"c": ["a", "b", "c"], "b": [True, False, True]})
     f = io.BytesIO()
     df.write_ipc(f)
     f.seek(0)
 
-    read_df = pl.read_ipc(f, columns=[1, 2], use_pyarrow=False)
+    read_df = pl.read_ipc(f, columns=[2, 1], use_pyarrow=False)
     assert expected.frame_equal(read_df)