fix explode empty df (#3405)

polars/polars-core/src/chunked_array/ops/explode.rs

@@ -237,17 +237,10 @@ impl ChunkExplode for ListChunked {
 
         // all empty
         if offsets[offsets.len() - 1] == 0 {
-            // Safety: empty dtype is correct
-            return unsafe {
-                Ok((
-                    Series::from_chunks_and_dtype_unchecked(
-                        self.name(),
-                        vec![values],
-                        &self.inner_dtype(),
-                    ),
-                    Buffer::from_slice(&[]),
-                ))
-            };
+            return Ok((
+                Series::new_empty(self.name(), &self.inner_dtype()),
+                Buffer::from_slice(&[]),
+            ));
         }
 
         if !offsets.is_empty() {

polars/polars-core/src/chunked_array/ops/mod.rs

@@ -124,7 +124,7 @@ pub trait ChunkTakeEvery<T> {
     fn take_every(&self, n: usize) -> ChunkedArray<T>;
 }
 
-/// Explode/ flatten a
+/// Explode/ flatten a List or Utf8 Series
 pub trait ChunkExplode {
     fn explode(&self) -> Result<Series> {
         self.explode_and_offsets().map(|t| t.0)

polars/polars-core/src/frame/explode.rs

@@ -29,8 +29,12 @@ pub struct MeltArgs {
 
 impl DataFrame {
     pub fn explode_impl(&self, mut columns: Vec<Series>) -> Result<DataFrame> {
+        let mut df = self.clone();
         if self.height() == 0 {
-            return Ok(self.clone());
+            for s in &columns {
+                df.with_column(s.explode()?)?;
+            }
+            return Ok(df);
         }
         columns.sort_by(|sa, sb| {
             self.check_name_to_idx(sa.name())
@@ -39,8 +43,6 @@ impl DataFrame {
                 .expect("cmp usize -> Ordering")
         });
 
-        let mut df = self.clone();
-
         // TODO: optimize this.
         // This is the slower easier option.
         // instead of filtering the whole dataframe first

py-polars/tests/test_explode.py

@@ -0,0 +1,8 @@
+import pyarrow as pa
+
+import polars as pl
+
+
+def test_explode_empty_df_3402() -> None:
+    df = pl.DataFrame({"a": pa.array([], type=pa.large_list(pa.int32()))})
+    assert df.explode("a").dtypes == [pl.Int32]