fix dataframe explode with empty lists (#3916)

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

@@ -208,6 +208,7 @@ pub(crate) fn offsets_to_indexes(offsets: &[i64], capacity: usize) -> Vec<IdxSiz
 
     let mut count = 0;
     let mut last_idx = 0;
+    let mut previous_empty = false;
     for offset in &offsets[1..] {
         while count < *offset {
             count += 1;
@@ -217,13 +218,19 @@ pub(crate) fn offsets_to_indexes(offsets: &[i64], capacity: usize) -> Vec<IdxSiz
         // we started iterating from 1, so there is always a previous offset
         // we take the pointer to the previous element and deref that to get
         // the previous offset
-        unsafe {
-            let previous_offset = *(offset as *const i64).offset(-1);
-            if previous_offset != *offset {
-                last_idx += 1;
-            }
+        let previous_offset = unsafe { *(offset as *const i64).offset(-1) };
+
+        if !previous_empty && (previous_offset != *offset) {
+            last_idx += 1;
+        } else {
+            count += 1;
+            idx.push(last_idx);
+            last_idx += 1;
         }
+        previous_empty = previous_offset == *offset;
     }
+    // last appended index.
+    let last_idx = idx[idx.len() - 1];
     for _ in 0..(capacity - count as usize) {
         idx.push(last_idx);
     }

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

@@ -367,9 +367,10 @@ mod test {
         let out = df.explode(["foo"])?;
         let expected = df![
             "foo" => [Some(1), Some(2), Some(3), None, Some(1), Some(1), Some(1)],
-            "B" => [1, 1, 1, 2, 2, 2, 3],
+            "B" => [1, 1, 1, 2, 3, 3, 3],
             "C" => [1, 1, 1, 1, 1, 1, 1],
         ]?;
+
         assert!(out.frame_equal_missing(&expected));
         Ok(())
     }

py-polars/tests/test_explode.py

@@ -11,3 +11,19 @@ def test_explode_empty_df_3402() -> None:
 def test_explode_empty_df_3460() -> None:
     df = pl.DataFrame({"a": pa.array([[]], type=pa.large_list(pa.int32()))})
     assert df.explode("a").dtypes == [pl.Int32]
+
+
+def test_explode_empty_df_3902() -> None:
+    df = pl.DataFrame(
+        {
+            "first": [1, 2, 3, 4, 5],
+            "second": [["a"], [], ["b", "c"], [], ["d", "f", "g"]],
+        }
+    )
+    expected = pl.DataFrame(
+        {
+            "first": [1, 2, 3, 3, 4, 5, 5, 5],
+            "second": ["a", None, "b", "c", None, "d", "f", "g"],
+        }
+    )
+    assert df.explode("second").frame_equal(expected)