diff --git a/src/stream/src/executor/backfill/arrangement_backfill.rs b/src/stream/src/executor/backfill/arrangement_backfill.rs
index 1cee14cd6eb52..2b51414df44b3 100644
--- a/src/stream/src/executor/backfill/arrangement_backfill.rs
+++ b/src/stream/src/executor/backfill/arrangement_backfill.rs
@@ -345,6 +345,7 @@ where
                         debug_assert!(builders.values().all(|b| b.is_empty()));
                         let (_, snapshot) = backfill_stream.into_inner();
                         let mut remaining_vnodes = vnodes.as_ref().clone();
+                        let mut yielded = false;
 
                         #[for_await]
                         for msg in snapshot {
@@ -369,14 +370,44 @@ where
                                     // for that vnode once we have read at least 1 record from it.
                                     let vnode_idx = vnode.to_index();
                                     if !remaining_vnodes.is_set(vnode_idx) {
-                                        let new_pos = row.project(&pk_in_output_indices);
-                                        assert_eq!(new_pos.len(), pk_in_output_indices.len());
-                                        backfill_state.update_progress(
-                                            vnode,
-                                            new_pos.to_owned_row(),
-                                            false,
-                                            0,
-                                        )?;
+                                        // FIXME(kwannoel): This is not needed.
+                                        // The main point of reading the snapshot here is by
+                                        // skip over long tombstone ranges, not to guarantee that snapshot read makes progress.
+                                        // Because skipping over long tombstone ranges is a transient step,
+                                        // once we skip over the initial run, we don't need to do it again.
+                                        //
+                                        // If we hit this code section, it means snapshot read is really slow.
+                                        // Even if we yielded a record here, the backfill progress will still be slow,
+                                        // since we only return one record per epoch.
+                                        // Consider barrier interval at 1s.
+                                        // And upstream has 1M records.
+                                        // So we take 1M seconds to finish backfill, which is 11.57 days.
+                                        // So that's basically the same as never finishing backfill.
+                                        if !yielded {
+                                            let builder = builders.get_mut(&vnode).unwrap();
+                                            if let Some(chunk) = builder.append_one_row(row) {
+                                                yield Message::Chunk(Self::handle_snapshot_chunk(
+                                                    chunk,
+                                                    vnode,
+                                                    &pk_in_output_indices,
+                                                    &mut backfill_state,
+                                                    &mut cur_barrier_snapshot_processed_rows,
+                                                    &mut total_snapshot_processed_rows,
+                                                    &self.output_indices,
+                                                )?);
+                                            }
+                                            yielded = true;
+                                        } else {
+                                            let new_pos = row.project(&pk_in_output_indices);
+                                            assert_eq!(new_pos.len(), pk_in_output_indices.len());
+                                            backfill_state.update_progress(
+                                                vnode,
+                                                new_pos.to_owned_row(),
+                                                false,
+                                                0,
+                                            )?;
+                                        }
+
                                         remaining_vnodes.set_bit(vnode_idx);
                                         if remaining_vnodes.is_empty() {
                                             break;