diff --git a/src/mito2/src/read/scan_region.rs b/src/mito2/src/read/scan_region.rs
index 2ff4850774ca..4fbffd325fbb 100644
--- a/src/mito2/src/read/scan_region.rs
+++ b/src/mito2/src/read/scan_region.rs
@@ -627,6 +627,12 @@ impl ScanPart {
         let end = current_range.1.max(part_range.1);
         self.time_range = Some((start, end));
     }
+
+    /// Returns true if the we can split the part into multiple parts
+    /// and preserving order.
+    pub(crate) fn can_split_preserve_order(&self) -> bool {
+        self.memtables.is_empty() && self.file_ranges.len() == 1 && self.file_ranges[0].len() > 1
+    }
 }
 
 /// A trait to collect file ranges to scan.
diff --git a/src/mito2/src/read/seq_scan.rs b/src/mito2/src/read/seq_scan.rs
index 51ac362dcf43..eabaec6bced3 100644
--- a/src/mito2/src/read/seq_scan.rs
+++ b/src/mito2/src/read/seq_scan.rs
@@ -334,10 +334,9 @@ impl SeqDistributor {
 
         let parallelism = parallelism.max(1);
         let parts = group_parts_by_range(self.parts);
-        let parts = maybe_merge_parts(parts, parallelism);
-        // TODO(yingwen): Split parts.
-
-        parts
+        let parts = maybe_split_parts(parts, parallelism);
+        // Ensures it doesn't returns parts more than `parallelism`.
+        maybe_merge_parts(parts, parallelism)
     }
 }
 
@@ -420,6 +419,62 @@ fn maybe_merge_parts(mut parts: Vec<ScanPart>, parallelism: usize) -> Vec<ScanPa
     parts
 }
 
+/// Splits parts by parallelism.
+/// It splits a part if it only scans one file and doesn't scan any memtable.
+fn maybe_split_parts(mut parts: Vec<ScanPart>, parallelism: usize) -> Vec<ScanPart> {
+    assert!(parallelism > 0);
+    if parts.len() >= parallelism {
+        // No need to split parts.
+        return parts;
+    }
+
+    let has_part_to_split = parts.iter().any(|part| part.can_split_preserve_order());
+    if !has_part_to_split {
+        // No proper parts to scan.
+        return parts;
+    }
+
+    // Sorts parts by the number of ranges in the first file.
+    parts.sort_unstable_by(|a, b| {
+        let a_len = a.file_ranges.first().map(|file| file.len()).unwrap_or(0);
+        let b_len = b.file_ranges.first().map(|file| file.len()).unwrap_or(0);
+        a_len.cmp(&b_len).reverse()
+    });
+    let num_parts_to_split = parallelism - parts.len();
+    let mut output_parts = Vec::with_capacity(parallelism);
+    // Split parts up to num_parts_to_split.
+    for part in parts.iter_mut() {
+        if !part.can_split_preserve_order() {
+            continue;
+        }
+        // Safety: `can_split_preserve_order()` ensures file_ranges.len() == 1.
+        // Splits part into `num_parts_to_split + 1` new parts if possible.
+        let target_part_num = num_parts_to_split + 1;
+        let ranges_per_part = (part.file_ranges[0].len() + target_part_num - 1) / target_part_num;
+        // `can_split_preserve_order()` ensures part.file_ranges[0].len() > 1.
+        assert!(ranges_per_part > 0);
+        for ranges in part.file_ranges[0].chunks(ranges_per_part) {
+            let new_part = ScanPart {
+                memtables: Vec::new(),
+                file_ranges: smallvec![ranges.to_vec()],
+                time_range: part.time_range,
+            };
+            output_parts.push(new_part);
+        }
+        // Replace the current part with the last output part as we will put the current part
+        // into the output parts later.
+        *part = output_parts.pop().unwrap();
+        if output_parts.len() >= num_parts_to_split {
+            // We already split enough parts.
+            break;
+        }
+    }
+    // Put the remaining parts into the output parts.
+    output_parts.append(&mut parts);
+
+    output_parts
+}
+
 #[cfg(test)]
 mod tests {
     use std::sync::Arc;