Add OderBy output stage spill (#7759)

Summary: Add a new type of spiller named `kOrderByOutput`. It is used to spill data during the output processing stage of `OrderBy` operator. The spiller uses a newly added API that supports spill data by row pointers. Pull Request resolved: #7759 Reviewed By: tanjialiang Differential Revision: D51947152 Pulled By: xiaoxmeng fbshipit-source-id: 10be66d8d4fa570b8a377d0ee7dac7bb18982841
facebookincubator · Dec 7, 2023 · ddc3471 · ddc3471
1 parent 61ab60d
commit ddc3471
Show file tree

Hide file tree

Showing 9 changed files with 251 additions and 81 deletions.
diff --git a/velox/exec/OrderBy.cpp b/velox/exec/OrderBy.cpp
@@ -78,24 +78,11 @@ void OrderBy::reclaim(
     memory::MemoryReclaimer::Stats& stats) {
   VELOX_CHECK(canReclaim());
   VELOX_CHECK(!nonReclaimableSection_);
-  auto* driver = operatorCtx_->driver();
 
-  // NOTE: an order by operator is reclaimable if it hasn't started output
-  // processing and is not under non-reclaimable execution section.
-  if (noMoreInput_) {
-    // TODO: reduce the log frequency if it is too verbose.
-    ++stats.numNonReclaimableAttempts;
-    LOG(WARNING)
-        << "Can't reclaim from order by operator which has started producing output: "
-        << pool()->name()
-        << ", usage: " << succinctBytes(pool()->currentBytes())
-        << ", reservation: " << succinctBytes(pool()->reservedBytes());
-    return;
-  }
-
-  // TODO: support fine-grain disk spilling based on 'targetBytes' after having
-  // row container memory compaction support later.
+  // TODO: support fine-grain disk spilling based on 'targetBytes' after
+  // having row container memory compaction support later.
   sortBuffer_->spill();
+
   // Release the minimum reserved memory.
   pool()->release();
 }

diff --git a/velox/exec/SortBuffer.cpp b/velox/exec/SortBuffer.cpp
@@ -16,7 +16,6 @@
 
 #include "SortBuffer.h"
 #include "velox/exec/MemoryReclaimer.h"
-#include "velox/vector/BaseVector.h"
 
 namespace facebook::velox::exec {
 
@@ -133,11 +132,7 @@ void SortBuffer::noMoreInput() {
     // for now.
     spill();
 
-    // Finish spill, and we shouldn't get any rows from non-spilled partition as
-    // there is only one hash partition for SortBuffer.
-    VELOX_CHECK_NULL(spillMerger_);
-    auto spillPartition = spiller_->finishSpill();
-    spillMerger_ = spillPartition.createOrderedReader(pool());
+    finishSpill();
   }
 
   // Releases the unused memory reservation after procesing input.
@@ -170,24 +165,11 @@ void SortBuffer::spill() {
   }
   updateEstimatedOutputRowSize();
 
-  if (spiller_ == nullptr) {
-    spiller_ = std::make_unique<Spiller>(
-        Spiller::Type::kOrderBy,
-        data_.get(),
-        spillerStoreType_,
-        data_->keyTypes().size(),
-        sortCompareFlags_,
-        spillConfig_->getSpillDirPathCb,
-        spillConfig_->fileNamePrefix,
-        spillConfig_->writeBufferSize,
-        spillConfig_->compressionKind,
-        memory::spillMemoryPool(),
-        spillConfig_->executor);
-    VELOX_CHECK_EQ(spiller_->state().maxPartitions(), 1);
+  if (sortedRows_.empty()) {
+    spillInput();
+  } else {
+    spillOutput();
   }
-
-  spiller_->spill();
-  data_->clear();
 }
 
 std::optional<uint64_t> SortBuffer::estimateOutputRowSize() const {
@@ -278,6 +260,52 @@ void SortBuffer::updateEstimatedOutputRowSize() {
   }
 }
 
+void SortBuffer::spillInput() {
+  if (spiller_ == nullptr) {
+    VELOX_CHECK(!noMoreInput_);
+    spiller_ = std::make_unique<Spiller>(
+        Spiller::Type::kOrderByInput,
+        data_.get(),
+        spillerStoreType_,
+        data_->keyTypes().size(),
+        sortCompareFlags_,
+        spillConfig_->getSpillDirPathCb,
+        spillConfig_->fileNamePrefix,
+        spillConfig_->writeBufferSize,
+        spillConfig_->compressionKind,
+        memory::spillMemoryPool(),
+        spillConfig_->executor);
+  }
+  spiller_->spill();
+  data_->clear();
+}
+
+void SortBuffer::spillOutput() {
+  if (spiller_ != nullptr) {
+    // Already spilled.
+    return;
+  }
+
+  spiller_ = std::make_unique<Spiller>(
+      Spiller::Type::kOrderByOutput,
+      data_.get(),
+      spillerStoreType_,
+      spillConfig_->getSpillDirPathCb,
+      spillConfig_->fileNamePrefix,
+      spillConfig_->writeBufferSize,
+      spillConfig_->compressionKind,
+      memory::spillMemoryPool(),
+      spillConfig_->executor);
+  auto spillRows = std::vector<char*>(
+      sortedRows_.begin() + numOutputRows_, sortedRows_.end());
+  spiller_->spill(spillRows);
+  data_->clear();
+  sortedRows_.clear();
+  // Finish right after spilling as the output spiller only spills at most
+  // once.
+  finishSpill();
+}
+
 void SortBuffer::prepareOutput(uint32_t maxOutputRows) {
   VELOX_CHECK_GT(maxOutputRows, 0);
   VELOX_CHECK_GT(numInputRows_, numOutputRows_);
@@ -364,4 +392,10 @@ void SortBuffer::getOutputWithSpill() {
   numOutputRows_ += output_->size();
 }
 
+void SortBuffer::finishSpill() {
+  VELOX_CHECK_NULL(spillMerger_);
+  auto spillPartition = spiller_->finishSpill();
+  spillMerger_ = spillPartition.createOrderedReader(pool());
+}
+
 } // namespace facebook::velox::exec
diff --git a/velox/exec/SortBuffer.h b/velox/exec/SortBuffer.h
@@ -80,6 +80,13 @@ class SortBuffer {
   void prepareOutput(uint32_t maxOutputRows);
   void getOutputWithoutSpill();
   void getOutputWithSpill();
+  // Spill during input stage.
+  void spillInput();
+  // Spill during output stage.
+  void spillOutput();
+  // Finish spill, and we shouldn't get any rows from non-spilled partition as
+  // there is only one hash partition for SortBuffer.
+  void finishSpill();
 
   const RowTypePtr input_;
   const std::vector<CompareFlags> sortCompareFlags_;

diff --git a/velox/exec/SortWindowBuild.cpp b/velox/exec/SortWindowBuild.cpp
@@ -140,7 +140,7 @@ void SortWindowBuild::setupSpiller() {
 
   spiller_ = std::make_unique<Spiller>(
       // TODO Replace Spiller::Type::kOrderBy.
-      Spiller::Type::kOrderBy,
+      Spiller::Type::kOrderByInput,
       data_.get(),
       inputType_,
       spillCompareFlags_.size(),

diff --git a/velox/exec/Spiller.cpp b/velox/exec/Spiller.cpp
@@ -61,7 +61,7 @@ Spiller::Spiller(
           executor,
           writeFileOptions) {
   VELOX_CHECK(
-      type_ == Type::kOrderBy || type_ == Type::kAggregateInput,
+      type_ == Type::kOrderByInput || type_ == Type::kAggregateInput,
       "Unexpected spiller type: {}",
       typeName(type_));
   VELOX_CHECK_EQ(state_.maxPartitions(), 1);
@@ -94,9 +94,8 @@ Spiller::Spiller(
           pool,
           executor,
           writeFileOptions) {
-  VELOX_CHECK_EQ(
-      type,
-      Type::kAggregateOutput,
+  VELOX_CHECK(
+      type_ == Type::kAggregateOutput || type_ == Type::kOrderByOutput,
       "Unexpected spiller type: {}",
       typeName(type_));
   VELOX_CHECK_EQ(state_.maxPartitions(), 1);
@@ -447,9 +446,11 @@ void Spiller::runSpill() {
     VELOX_CHECK_EQ(numWritten, run.rows.size());
     run.clear();
     // When a sorted run ends, we start with a new file next time. For
-    // aggregation output spiller, we expect only one spill call to spill all
-    // the rows starting from the specified row offset.
-    if (needSort() || (type_ == Spiller::Type::kAggregateOutput)) {
+    // aggregation output / orderby output spiller, we expect only one spill
+    // call to spill all the rows starting from the specified row offset.
+    if (needSort() ||
+        (type_ == Spiller::Type::kAggregateOutput ||
+         type_ == Spiller::Type::kOrderByOutput)) {
       state_.finishFile(partition);
     }
   }
@@ -467,7 +468,7 @@ void Spiller::updateSpillSortTime(uint64_t timeUs) {
 
 bool Spiller::needSort() const {
   return type_ != Type::kHashJoinProbe && type_ != Type::kHashJoinBuild &&
-      type_ != Type::kAggregateOutput;
+      type_ != Type::kAggregateOutput && type_ != Type::kOrderByOutput;
 }
 
 void Spiller::spill() {
@@ -483,25 +484,41 @@ void Spiller::spill(const RowContainerIterator* startRowIter) {
   CHECK_NOT_FINALIZED();
   VELOX_CHECK_NE(type_, Type::kHashJoinProbe);
 
-  // Marks all the partitions have been spilled as we don't support fine-grained
-  // spilling as for now.
-  for (auto partition = 0; partition < state_.maxPartitions(); ++partition) {
-    if (!state_.isPartitionSpilled(partition)) {
-      state_.setPartitionSpilled(partition);
-    }
-  }
+  markAllPartitionsSpilled();
 
   fillSpillRuns(startRowIter);
   runSpill();
   checkEmptySpillRuns();
 }
 
+void Spiller::spill(std::vector<char*>& rows) {
+  CHECK_NOT_FINALIZED();
+  VELOX_CHECK_EQ(type_, Type::kOrderByOutput);
+  if (rows.empty()) {
+    return;
+  }
+
+  markAllPartitionsSpilled();
+
+  fillSpillRun(rows);
+  runSpill();
+  checkEmptySpillRuns();
+}
+
 void Spiller::checkEmptySpillRuns() const {
   for (const auto& spillRun : spillRuns_) {
     VELOX_CHECK(spillRun.rows.empty());
   }
 }
 
+void Spiller::markAllPartitionsSpilled() {
+  for (auto partition = 0; partition < state_.maxPartitions(); ++partition) {
+    if (!state_.isPartitionSpilled(partition)) {
+      state_.setPartitionSpilled(partition);
+    }
+  }
+}
+
 void Spiller::spill(uint32_t partition, const RowVectorPtr& spillVector) {
   CHECK_NOT_FINALIZED();
   VELOX_CHECK(
@@ -598,6 +615,21 @@ void Spiller::fillSpillRuns(const RowContainerIterator* startRowIter) {
   updateSpillFillTime(execTimeUs);
 }
 
+void Spiller::fillSpillRun(std::vector<char*>& rows) {
+  VELOX_CHECK_EQ(bits_.numPartitions(), 1);
+  checkEmptySpillRuns();
+  uint64_t execTimeUs{0};
+  {
+    MicrosecondTimer timer(&execTimeUs);
+    spillRuns_[0].rows =
+        SpillRows(rows.begin(), rows.end(), spillRuns_[0].rows.get_allocator());
+    for (const auto* row : rows) {
+      spillRuns_[0].numBytes += container_->rowSize(row);
+    }
+  }
+  updateSpillFillTime(execTimeUs);
+}
+
 std::string Spiller::toString() const {
   return fmt::format(
       "{}\t{}\tMAX_PARTITIONS:{}\tFINALIZED:{}",
@@ -610,8 +642,10 @@ std::string Spiller::toString() const {
 // static
 std::string Spiller::typeName(Type type) {
   switch (type) {
-    case Type::kOrderBy:
-      return "ORDER_BY";
+    case Type::kOrderByInput:
+      return "ORDER_BY_INPUT";
+    case Type::kOrderByOutput:
+      return "ORDER_BY_OUTPUT";
     case Type::kHashJoinBuild:
       return "HASH_JOIN_BUILD";
     case Type::kHashJoinProbe:

diff --git a/velox/exec/Spiller.h b/velox/exec/Spiller.h
@@ -35,11 +35,14 @@ class Spiller {
     kHashJoinBuild = 2,
     // Used for hash join probe.
     kHashJoinProbe = 3,
-    // Used for order by.
-    kOrderBy = 4,
+    // Used for order by input processing stage.
+    kOrderByInput = 4,
+    // Used for order by output processing stage.
+    kOrderByOutput = 5,
     // Number of spiller types.
-    kNumTypes = 5,
+    kNumTypes = 6,
   };
+
   static std::string typeName(Type);
 
   using SpillRows = std::vector<char*, memory::StlAllocator<char*>>;
@@ -118,6 +121,12 @@ class Spiller {
   /// The caller needs to erase them from the row container.
   void spill(const RowContainerIterator& startRowIter);
 
+  /// Invoked to spill all the rows pointed by rows. This is used by
+  /// 'kOrderByOutput' spiller type to spill during the order by
+  /// output processing. Similarly, the spilled rows still stays in the row
+  /// container. The caller needs to erase them from the row container.
+  void spill(std::vector<char*>& rows);
+
   /// Append 'spillVector' into the spill file of given 'partition'. It is now
   /// only used by the spilling operator which doesn't need data sort, such as
   /// hash join build and hash join probe.
@@ -274,11 +283,19 @@ class Spiller {
 
   void checkEmptySpillRuns() const;
 
+  // Marks all the partitions have been spilled as we don't support
+  // fine-grained spilling as for now.
+  void markAllPartitionsSpilled();
+
   // Prepares spill runs for the spillable data from all the hash partitions.
   // If 'startRowIter' is not null, we prepare runs starting from the offset
   // pointed by 'startRowIter'.
   void fillSpillRuns(const RowContainerIterator* startRowIter = nullptr);
 
+  // Prepares spill run of a single partition for the spillable data from the
+  // rows.
+  void fillSpillRun(std::vector<char*>& rows);
+
   // Writes out all the rows collected in spillRuns_.
   void runSpill();
 

diff --git a/velox/exec/TopNRowNumber.cpp b/velox/exec/TopNRowNumber.cpp
@@ -737,7 +737,7 @@ void TopNRowNumber::setupSpiller() {
 
   spiller_ = std::make_unique<Spiller>(
       // TODO Replace Spiller::Type::kOrderBy.
-      Spiller::Type::kOrderBy,
+      Spiller::Type::kOrderByInput,
       data_.get(),
       inputType_,
       spillCompareFlags_.size(),

diff --git a/velox/exec/tests/OrderByTest.cpp b/velox/exec/tests/OrderByTest.cpp
@@ -1082,15 +1082,10 @@ DEBUG_ONLY_TEST_F(OrderByTest, reclaimDuringOutputProcessing) {
 
     if (enableSpilling) {
       ASSERT_GT(reclaimableBytes, 0);
-      const auto usedMemoryBytes = op->pool()->currentBytes();
-      reclaimAndRestoreCapacity(
-          op,
-          folly::Random::oneIn(2) ? 0 : folly::Random::rand32(rng_),
-          reclaimerStats_);
-      ASSERT_GT(reclaimerStats_.reclaimedBytes, 0);
+      reclaimerStats_.reset();
+      reclaimAndRestoreCapacity(op, reclaimableBytes, reclaimerStats_);
+      ASSERT_EQ(reclaimerStats_.reclaimedBytes, reclaimableBytes);
       ASSERT_GT(reclaimerStats_.reclaimExecTimeUs, 0);
-      // No reclaim as the operator has started output processing.
-      ASSERT_EQ(usedMemoryBytes, op->pool()->currentBytes());
     } else {
       ASSERT_EQ(reclaimableBytes, 0);
       VELOX_ASSERT_THROW(
@@ -1108,7 +1103,7 @@ DEBUG_ONLY_TEST_F(OrderByTest, reclaimDuringOutputProcessing) {
     ASSERT_EQ(stats[0].operatorStats[1].spilledPartitions, 0);
     OperatorTestBase::deleteTaskAndCheckSpillDirectory(task);
   }
-  ASSERT_EQ(reclaimerStats_.numNonReclaimableAttempts, 1);
+  ASSERT_EQ(reclaimerStats_.numNonReclaimableAttempts, 0);
 }
 
 DEBUG_ONLY_TEST_F(OrderByTest, abortDuringOutputProcessing) {