c

crates/polars-plan/src/plans/conversion/dsl_to_ir.rs

@@ -380,15 +380,25 @@ pub fn to_alp_impl(lp: DslPlan, ctxt: &mut DslConversionContext) -> PolarsResult
             expr,
             input,
             options,
-        } if
+        } if {
             // This is a hack so that we don't break `list.eval(pl.element().filter())`. That expression
             // hits this codepath with `col("").filter(..)` when it goes through `run_on_group_by_engine`
             // and for some reason doesn't give a correct result if we do this rewrite.
             //
             // We detect that we are called by `run_on_group_by_engine` because it calls to here with
             // nearly all optimizations turned off, so we just picked the `PREDICATE_PUSHDOWN` to check.
             ctxt.opt_flags.contains(OptFlags::PREDICATE_PUSHDOWN)
-            && matches!(expr.as_slice(), &[Expr::Filter { .. }]) =>
+        } && {
+            let mut rewrite = false;
+
+            if let [Expr::Filter { input, .. }] = &expr[..] {
+                if has_expr(input.as_ref(), |e| matches!(e, Expr::Wildcard)) {
+                    rewrite = true
+                }
+            }
+
+            rewrite
+        } =>
         {
             let Expr::Filter {
                 input: filter_input,

py-polars/tests/unit/expr/test_exprs.py

@@ -680,6 +680,11 @@ def test_filter_all() -> None:
         pl.DataFrame({"a": [2, 5], "b": ["q", "t"], "p": False}),
     )
 
+    # Ensure no regression in the non-wildcard case, as the predicate may refer
+    # to columns that are not part of the `select()`
+    q = df.lazy().select(pl.col("a").filter(~pl.col("p")))
+    assert_frame_equal(q.collect(), pl.DataFrame({"a": [2, 5]}))
+
     q = df.lazy().select((pl.all().reverse()).filter(~pl.col("p")))
     assert r'FILTER col("p").not()' in q.explain()