Fix: External sort failing on StringView due to shared buffers

datafusion/core/tests/memory_limit/mod.rs

@@ -23,9 +23,10 @@ use std::sync::{Arc, LazyLock};
 
 #[cfg(feature = "extended_tests")]
 mod memory_limit_validation;
-use arrow::array::{ArrayRef, DictionaryArray, RecordBatch};
+use arrow::array::{ArrayRef, DictionaryArray, Int32Array, RecordBatch, StringViewArray};
 use arrow::compute::SortOptions;
 use arrow::datatypes::{Int32Type, SchemaRef};
+use arrow_schema::{DataType, Field, Schema};
 use datafusion::assert_batches_eq;
 use datafusion::datasource::memory::MemorySourceConfig;
 use datafusion::datasource::source::DataSourceExec;
@@ -41,14 +42,15 @@ use datafusion_catalog::streaming::StreamingTable;
 use datafusion_catalog::Session;
 use datafusion_common::{assert_contains, Result};
 use datafusion_execution::memory_pool::{
-    GreedyMemoryPool, MemoryPool, TrackConsumersPool,
+    FairSpillPool, GreedyMemoryPool, MemoryPool, TrackConsumersPool,
 };
 use datafusion_execution::TaskContext;
 use datafusion_expr::{Expr, TableType};
 use datafusion_physical_expr::{LexOrdering, PhysicalSortExpr};
 use datafusion_physical_optimizer::join_selection::JoinSelection;
 use datafusion_physical_optimizer::PhysicalOptimizerRule;
 use datafusion_physical_plan::spill::get_record_batch_memory_size;
+use rand::Rng;
 use test_utils::AccessLogGenerator;
 
 use async_trait::async_trait;
@@ -403,6 +405,69 @@ async fn oom_with_tracked_consumer_pool() {
         .await
 }
 
+/// For regression case: if spilled `StringViewArray`'s buffer will be referenced by
+/// other batches which are also need to be spilled, then the spill writer will
+/// repeatedly write out the same buffer, and after reading back, each batch's size
+/// will explode.
+///
+/// This test setup will cause 10 spills, each spill will sort around 20 batches.
+/// If there is memory explosion for spilled record batch, this test will fail.
+#[tokio::test]
+async fn test_stringview_external_sort() {
+    let mut rng = rand::thread_rng();
+    let array_length = 1000;
+    let num_batches = 200;
+    // Batches contain two columns: random 100-byte string, and random i32
+    let mut batches = Vec::with_capacity(num_batches);
+
+    for _ in 0..num_batches {
+        let strings: Vec<String> = (0..array_length)
+            .map(|_| {
+                (0..100)
+                    .map(|_| rng.gen_range(0..=u8::MAX) as char)
+                    .collect()
+            })
+            .collect();
+
+        let string_array = StringViewArray::from(strings);
+        let array_ref: ArrayRef = Arc::new(string_array);
+
+        let random_numbers: Vec<i32> =
+            (0..array_length).map(|_| rng.gen_range(0..=1000)).collect();
+        let int_array = Int32Array::from(random_numbers);
+        let int_array_ref: ArrayRef = Arc::new(int_array);
+
+        let batch = RecordBatch::try_new(
+            Arc::new(Schema::new(vec![
+                Field::new("strings", DataType::Utf8View, false),
+                Field::new("random_numbers", DataType::Int32, false),
+            ])),
+            vec![array_ref, int_array_ref],
+        )
+        .unwrap();
+        batches.push(batch);
+    }
+
+    // Run a sql query that sorts the batches by the int column
+    let schema = batches[0].schema();
+    let table = MemTable::try_new(schema, vec![batches]).unwrap();
+    let builder = RuntimeEnvBuilder::new()
+        .with_memory_pool(Arc::new(FairSpillPool::new(60 * 1024 * 1024)));
+    let runtime = builder.build_arc().unwrap();
+
+    let config = SessionConfig::new().with_sort_spill_reservation_bytes(40 * 1024 * 1024);
+
+    let ctx = SessionContext::new_with_config_rt(config, runtime);
+    ctx.register_table("t", Arc::new(table)).unwrap();
+
+    let df = ctx
+        .sql("explain analyze SELECT * FROM t ORDER BY random_numbers")
+        .await
+        .unwrap();
+
+    let _ = df.collect().await.expect("Query execution failed");
+}
+
 /// Run the query with the specified memory limit,
 /// and verifies the expected errors are returned
 #[derive(Clone, Debug)]

datafusion/physical-plan/src/sorts/sort.rs

@@ -44,7 +44,9 @@ use crate::{
     Statistics,
 };
 
-use arrow::array::{Array, RecordBatch, RecordBatchOptions, UInt32Array};
+use arrow::array::{
+    Array, RecordBatch, RecordBatchOptions, StringViewArray, UInt32Array,
+};
 use arrow::compute::{concat_batches, lexsort_to_indices, take_arrays, SortColumn};
 use arrow::datatypes::{DataType, SchemaRef};
 use arrow::row::{RowConverter, SortField};
@@ -300,6 +302,7 @@ impl ExternalSorter {
         if input.num_rows() == 0 {
             return Ok(());
         }
+
         self.reserve_memory_for_merge()?;
 
         let size = get_reserved_byte_for_record_batch(&input);
@@ -397,6 +400,8 @@ impl ExternalSorter {
             return Ok(0);
         }
 
+        self.organize_stringview_arrays()?;
+
         debug!("Spilling sort data of ExternalSorter to disk whilst inserting");
 
         let spill_file = self.runtime.disk_manager.create_tmp_file("Sorting")?;
@@ -414,6 +419,66 @@ impl ExternalSorter {
         Ok(used)
     }
 
+    /// Reconstruct `self.in_mem_batches` to organize the payload buffers of each
+    /// `StringViewArray` in sequential order by calling `gc()` on them.
+    ///
+    /// # Rationale
+    /// After (merge-based) sorting, all batches will be sorted into a single run,
+    /// but physically this sorted run is chunked into many small batches. For
+    /// `StringViewArray`s inside each sorted run, their inner buffers are not
+    /// re-constructed by default, leading to non-sequential payload locations
+    /// (permutated by `interleave()` Arrow kernel). A single payload buffer might
+    /// be shared by multiple `RecordBatch`es.
+    /// When writing each batch to disk, the writer has to write all referenced buffers,
+    /// because they have to be read back one by one to reduce memory usage. This
+    /// causes extra disk reads and writes, and potentially execution failure.
+    ///
+    /// # Example
+    /// Before sorting:
+    /// batch1 -> buffer1
+    /// batch2 -> buffer2
+    ///
+    /// sorted_batch1 -> buffer1
+    ///               -> buffer2
+    /// sorted_batch2 -> buffer1
+    ///               -> buffer2
+    ///
+    /// Then when spilling each batch, the writer has to write all referenced buffers
+    /// repeatedly.
+    fn organize_stringview_arrays(&mut self) -> Result<()> {
+        let mut organized_batches = Vec::with_capacity(self.in_mem_batches.len());
+
+        for batch in self.in_mem_batches.drain(..) {
+            let mut new_columns: Vec<Arc<dyn Array>> =
+                Vec::with_capacity(batch.num_columns());
+
+            let mut arr_mutated = false;
+            for array in batch.columns() {
+                if let Some(string_view_array) =
+                    array.as_any().downcast_ref::<StringViewArray>()
+                {
+                    let new_array = string_view_array.gc();
+                    new_columns.push(Arc::new(new_array));
+                    arr_mutated = true;
+                } else {
+                    new_columns.push(Arc::clone(array));
+                }
+            }
+
+            let organized_batch = if arr_mutated {
+                RecordBatch::try_new(batch.schema(), new_columns)?
+            } else {
+                batch
+            };
+
+            organized_batches.push(organized_batch);
+        }
+
+        self.in_mem_batches = organized_batches;
+
+        Ok(())
+    }
+
     /// Sorts the in_mem_batches in place
     ///
     /// Sorting may have freed memory, especially if fetch is `Some`. If
@@ -446,21 +511,12 @@ impl ExternalSorter {
                 None => {
                     let sorted_size = get_reserved_byte_for_record_batch(&batch);
                     if self.reservation.try_grow(sorted_size).is_err() {
-                        // Directly write in_mem_batches as well as all the remaining batches in
-                        // sorted_stream to disk. Further batches fetched from `sorted_stream` will
-                        // be handled by the `Some(writer)` matching arm.
-                        let spill_file =
-                            self.runtime.disk_manager.create_tmp_file("Sorting")?;
-                        let mut writer = IPCWriter::new(spill_file.path(), &self.schema)?;
-                        // Flush everything in memory to the spill file
-                        for batch in self.in_mem_batches.drain(..) {
-                            writer.write(&batch)?;
-                        }
-                        // as well as the newly sorted batch
-                        writer.write(&batch)?;
-                        spill_writer = Some(writer);
+                        // Although the reservation is not enough, the batch is
+                        // already in memory, so it's okay to combine it with previously
+                        // sorted batches, and spill together.
+                        self.in_mem_batches.push(batch);
+                        self.spill().await?;
                         self.reservation.free();
-                        self.spills.push(spill_file);
                     } else {
                         self.in_mem_batches.push(batch);
                         self.in_mem_batches_sorted = true;
@@ -1425,7 +1481,7 @@ mod tests {
         // Processing 840 KB of data using 400 KB of memory requires at least 2 spills
         // It will spill roughly 18000 rows and 800 KBytes.
         // We leave a little wiggle room for the actual numbers.
-        assert!((2..=10).contains(&spill_count));
+        assert!((12..=18).contains(&spill_count));
         assert!((15000..=20000).contains(&spilled_rows));
         assert!((700000..=900000).contains(&spilled_bytes));