groupby apply: use inner type to infer dtype (#3955)

polars/polars-lazy/src/physical_plan/expressions/apply.rs

@@ -98,8 +98,16 @@ impl PhysicalExpr for ApplyExpr {
                     // collection of empty list leads to a null dtype
                     // see: #3687
                     if s.len() == 0 {
-                        let s = self.function.call_udf(&mut [s.clone()])?;
-                        let ca = ListChunked::full(s.name(), &s, 0);
+                        // create input for the function to determine the output dtype
+                        // see #3946
+                        let agg = ac.aggregated();
+                        let agg = agg.list().unwrap();
+                        let input_dtype = agg.inner_dtype();
+
+                        let input = Series::full_null("", 0, &input_dtype);
+
+                        let output = self.function.call_udf(&mut [input])?;
+                        let ca = ListChunked::full(ac.series().name(), &output, 0);
                         return Ok(self.finish_apply_groups(ac, ca));
                     }
 

polars/tests/it/lazy/expressions/apply.rs

@@ -82,3 +82,26 @@ fn test_expand_list() -> Result<()> {
 
     Ok(())
 }
+
+#[test]
+fn test_apply_groups_empty() -> Result<()> {
+    let df = df![
+        "id" => [1, 1],
+        "hi" => ["here", "here"]
+    ]?;
+
+    let out = df
+        .lazy()
+        .filter(col("id").eq(lit(2)))
+        .groupby([col("id")])
+        .agg([col("hi").drop_nulls().unique()])
+        .collect()?;
+
+    assert_eq!(
+        out.dtypes(),
+        &[DataType::Int32, DataType::List(Box::new(DataType::Utf8))]
+    );
+    assert_eq!(out.shape(), (0, 2));
+
+    Ok(())
+}