fix: Handle Parquet projection pushdown with only row index (#18520)

crates/polars-io/src/parquet/read/read_impl.rs

@@ -5,6 +5,7 @@ use std::ops::{Deref, Range};
 use arrow::array::BooleanArray;
 use arrow::bitmap::MutableBitmap;
 use arrow::datatypes::ArrowSchemaRef;
+use polars_core::chunked_array::builder::NullChunkedBuilder;
 use polars_core::prelude::*;
 use polars_core::utils::{accumulate_dataframes_vertical, split_df};
 use polars_core::POOL;
@@ -151,6 +152,20 @@ fn rg_to_dfs(
     use_statistics: bool,
     hive_partition_columns: Option<&[Series]>,
 ) -> PolarsResult<Vec<DataFrame>> {
+    // If we are only interested in the row_index, we take a little special path here.
+    if projection.is_empty() {
+        if let Some(row_index) = row_index {
+            let placeholder =
+                NullChunkedBuilder::new(PlSmallStr::from_static("__PL_TMP"), slice.1).finish();
+            return Ok(vec![DataFrame::new(vec![placeholder.into_series()])?
+                .with_row_index(
+                    row_index.name.clone(),
+                    Some(row_index.offset + IdxSize::try_from(slice.0).unwrap()),
+                )?
+                .select(std::iter::once(row_index.name))?]);
+        }
+    }
+
     use ParallelStrategy as S;
 
     if parallel == S::Prefiltered {

py-polars/tests/unit/io/test_parquet.py

@@ -1868,3 +1868,38 @@ def test_empty_parquet() -> None:
 
     empty_from_pl = pl.read_parquet(f_pl)
     assert empty_from_pl.shape == (0, 0)
+
+
+@pytest.mark.parametrize(
+    "strategy",
+    ["columns", "row_groups", "prefiltered"],
+)
+@pytest.mark.write_disk
+def test_row_index_projection_pushdown_18463(
+    tmp_path: Path, strategy: pl.ParallelStrategy
+) -> None:
+    tmp_path.mkdir(exist_ok=True)
+    f = tmp_path / "test.parquet"
+
+    pl.DataFrame({"A": [1, 4], "B": [2, 5]}).write_parquet(f)
+
+    df = pl.scan_parquet(f, parallel=strategy).with_row_index()
+
+    assert_frame_equal(df.select("index").collect(), df.collect().select("index"))
+
+    df = pl.scan_parquet(f, parallel=strategy).with_row_index("other_idx_name")
+
+    assert_frame_equal(
+        df.select("other_idx_name").collect(), df.collect().select("other_idx_name")
+    )
+
+    df = pl.scan_parquet(f, parallel=strategy).with_row_index(offset=42)
+
+    assert_frame_equal(df.select("index").collect(), df.collect().select("index"))
+
+    df = pl.scan_parquet(f, parallel=strategy).with_row_index()
+
+    assert_frame_equal(
+        df.select("index").slice(1, 1).collect(),
+        df.collect().select("index").slice(1, 1),
+    )