fix: Fix invalid len due to projection pushdown selection of scalar

crates/polars-plan/src/plans/optimizer/projection_pushdown/projection.rs

@@ -67,25 +67,66 @@ pub(super) fn process_projection(
         let expr = if input_schema.is_empty() {
             // If the input schema is empty, we should just project
             // ourselves
-            exprs[0].node()
+            Some(exprs[0].node())
         } else {
-            // simply select the last column
-            // NOTE: the first can be the inserted index column, so that might not work
-            let (first_name, _) = input_schema.try_get_at_index(input_schema.len() - 1)?;
-            let expr = expr_arena.add(AExpr::Column(first_name.clone()));
-            if !acc_projections.is_empty() {
-                check_double_projection(
-                    &exprs[0],
-                    expr_arena,
-                    &mut acc_projections,
-                    &mut projected_names,
-                );
+            // Select the last column projection.
+            let mut name = None;
+            for (_, plan) in (&*lp_arena).iter(input) {
+                match plan {
+                    IR::Select { expr: exprs, .. } | IR::HStack { exprs, .. } => {
+                        for e in exprs {
+                            if !e.is_scalar(expr_arena) {
+                                name = Some(e.output_name());
+                                break;
+                            }
+                        }
+                    },
+                    IR::Scan {
+                        file_info,
+                        output_schema,
+                        ..
+                    } => {
+                        let schema = output_schema.as_ref().unwrap_or(&file_info.schema);
+                        // NOTE: the first can be the inserted index column, so that might not work
+                        let (last_name, _) = schema.try_get_at_index(schema.len() - 1)?;
+                        name = Some(last_name);
+                        break;
+                    },
+                    IR::DataFrameScan {
+                        schema,
+                        output_schema,
+                        ..
+                    } => {
+                        // NOTE: the first can be the inserted index column, so that might not work
+                        let schema = output_schema.as_ref().unwrap_or(schema);
+                        let (last_name, _) = schema.try_get_at_index(schema.len() - 1)?;
+                        name = Some(last_name);
+                        break;
+                    },
+                    _ => {},
+                }
+            }
+
+            if let Some(name) = name {
+                let expr = expr_arena.add(AExpr::Column(name.clone()));
+                if !acc_projections.is_empty() {
+                    check_double_projection(
+                        &exprs[0],
+                        expr_arena,
+                        &mut acc_projections,
+                        &mut projected_names,
+                    );
+                }
+                Some(expr)
+            } else {
+                None
             }
-            expr
         };
-        add_expr_to_accumulated(expr, &mut acc_projections, &mut projected_names, expr_arena);
-        local_projection.push(exprs.pop().unwrap());
-        proj_pd.is_count_star = true;
+        if let Some(expr) = expr {
+            add_expr_to_accumulated(expr, &mut acc_projections, &mut projected_names, expr_arena);
+            local_projection.push(exprs.pop().unwrap());
+            proj_pd.is_count_star = true;
+        }
     } else {
         // A projection can consist of a chain of expressions followed by an alias.
         // We want to do the chain locally because it can have complicated side effects.

py-polars/tests/unit/test_scalar.py

@@ -47,3 +47,19 @@ def test_scalar_19957() -> None:
         "foo": [1, 1, 1],
         "bar": [1, 1, 1],
     }
+
+
+def test_scalar_len_20046() -> None:
+    df = pl.DataFrame({"a": [1, 2, 3]})
+
+    assert (
+        df.lazy()
+        .select(
+            pl.col("a"),
+            pl.lit(1),
+        )
+        .select(pl.len())
+        .collect()
+        .item()
+        == 3
+    )