Add clustered Iceberg writer mode.

Author: PingLiuPing

velox/connectors/hive/HiveConfig.cpp

@@ -273,4 +273,9 @@ std::string HiveConfig::schema(const config::ConfigBase* session) const {
       kSchema, config_->get<std::string>(kSchema, ""));
 }
 
+bool HiveConfig::fanoutEnabled(const config::ConfigBase* session) const {
+  return session->get<bool>(
+      kFanoutEnabledSession, config_->get<bool>(kFanoutEnabled, true));
+}
+
 } // namespace facebook::velox::connector::hive

velox/connectors/hive/HiveConfig.h

@@ -208,6 +208,12 @@ class HiveConfig {
   static constexpr const char* kSource = "source";
   static constexpr const char* kSchema = "schema";
 
+  /// Controls the writer mode, whether the fanout mode writer is enabled,
+  /// default value is true, setting to false means clustered mode.
+  /// Currently applies only to the Iceberg writer.
+  static constexpr const char* kFanoutEnabled = "fanout-enabled";
+  static constexpr const char* kFanoutEnabledSession = "fanout_enabled";
+
   InsertExistingPartitionsBehavior insertExistingPartitionsBehavior(
       const config::ConfigBase* session) const;
 
@@ -305,6 +311,9 @@ class HiveConfig {
   /// Schema of the query. Used for storage logging.
   std::string schema(const config::ConfigBase* session) const;
 
+  /// Return if fanout writer mode is enabled.
+  bool fanoutEnabled(const config::ConfigBase* session) const;
+
   HiveConfig(std::shared_ptr<const config::ConfigBase> config) {
     VELOX_CHECK_NOT_NULL(
         config, "Config is null for HiveConfig initialization");

velox/connectors/hive/iceberg/IcebergDataSink.cpp

@@ -30,6 +30,9 @@ namespace facebook::velox::connector::hive::iceberg {
 
 namespace {
 
+constexpr std::string_view kNotClusteredRowsErrorMsg =
+    "Incoming records violate the writer assumption that records are clustered by spec and \n by partition within each spec. Either cluster the incoming records or switch to fanout writers.\nEncountered records that belong to already closed files:\n";
+
 #define WRITER_NON_RECLAIMABLE_SECTION_GUARD(index)       \
   memory::NonReclaimableSectionGuard nonReclaimableGuard( \
       writerInfo_[(index)]->nonReclaimableSectionHolder.get())
@@ -213,7 +216,10 @@ IcebergDataSink::IcebergDataSink(
                     insertTableHandle->columnTransforms(),
                     hiveConfig->isPartitionPathAsLowerCase(
                         connectorQueryCtx->sessionProperties()))
-              : nullptr) {
+              : nullptr),
+      fanoutEnabled_(
+          hiveConfig_->fanoutEnabled(connectorQueryCtx_->sessionProperties())),
+      currentWriterId_(0) {
   if (isPartitioned()) {
     partitionData_.resize(maxOpenWriters_);
   }
@@ -332,8 +338,6 @@ std::vector<std::string> IcebergDataSink::commitMessage() const {
 }
 
 void IcebergDataSink::splitInputRowsAndEnsureWriters(RowVectorPtr input) {
-  VELOX_CHECK(isPartitioned());
-
   std::fill(partitionSizes_.begin(), partitionSizes_.end(), 0);
 
   const auto numRows = partitionIds_.size();
@@ -346,26 +350,7 @@ void IcebergDataSink::splitInputRowsAndEnsureWriters(RowVectorPtr input) {
     if (!partitionData_[index].empty()) {
       continue;
     }
-
-    std::vector<folly::dynamic> partitionValues(partitionChannels_.size());
-    auto icebergPartitionIdGenerator =
-        dynamic_cast<const IcebergPartitionIdGenerator*>(
-            partitionIdGenerator_.get());
-    VELOX_CHECK_NOT_NULL(icebergPartitionIdGenerator);
-    const RowVectorPtr transformedValues =
-        icebergPartitionIdGenerator->partitionValues();
-    for (auto i = 0; i < partitionChannels_.size(); ++i) {
-      auto block = transformedValues->childAt(i);
-      if (block->isNullAt(index)) {
-        partitionValues[i] = nullptr;
-      } else {
-        DecodedVector decoded(*block);
-        partitionValues[i] = VELOX_DYNAMIC_SCALAR_TYPE_DISPATCH(
-            extractPartitionValue, block->typeKind(), &decoded, index);
-      }
-    }
-
-    partitionData_[index] = partitionValues;
+    buildPartitionData(index);
   }
 
   for (auto i = 0; i < partitionSizes_.size(); ++i) {
@@ -376,6 +361,11 @@ void IcebergDataSink::splitInputRowsAndEnsureWriters(RowVectorPtr input) {
   }
 }
 
+void IcebergDataSink::computePartition(const RowVectorPtr& input) {
+  VELOX_CHECK(isPartitioned());
+  partitionIdGenerator_->run(input, partitionIds_);
+}
+
 void IcebergDataSink::appendData(RowVectorPtr input) {
   checkRunning();
   if (!isPartitioned()) {
@@ -384,22 +374,79 @@ void IcebergDataSink::appendData(RowVectorPtr input) {
     return;
   }
 
-  // Compute partition and bucket numbers.
-  computePartitionAndBucketIds(input);
+  computePartition(input);
 
-  splitInputRowsAndEnsureWriters(input);
+  if (fanoutEnabled_) {
+    splitInputRowsAndEnsureWriters(input);
 
-  for (auto index = 0; index < writers_.size(); ++index) {
-    const vector_size_t partitionSize = partitionSizes_[index];
-    if (partitionSize == 0) {
-      continue;
+    for (auto index = 0; index < writers_.size(); ++index) {
+      const vector_size_t partitionSize = partitionSizes_[index];
+      if (partitionSize == 0) {
+        continue;
+      }
+
+      const RowVectorPtr writerInput = partitionSize == input->size()
+          ? input
+          : exec::wrap(partitionSize, partitionRows_[index], input);
+      write(index, writerInput);
+    }
+  } else { // Clustered mode.
+    std::fill(partitionSizes_.begin(), partitionSizes_.end(), 0);
+    const auto numRows = input->size();
+    uint32_t index = 0;
+    for (auto row = 0; row < numRows; ++row) {
+      auto id = getIcebergWriterId(row);
+      index = ensureWriter(id);
+      if (currentWriterId_ != index) {
+        clusteredWrite(input, currentWriterId_);
+        closeWriter(currentWriterId_);
+        completedWriterIds_.insert(currentWriterId_);
+        VELOX_USER_CHECK_EQ(
+            completedWriterIds_.count(index),
+            0,
+            "{}",
+            kNotClusteredRowsErrorMsg);
+        currentWriterId_ = index;
+      }
+      updatePartitionRows(index, numRows, row);
+      buildPartitionData(index);
     }
+    clusteredWrite(input, index);
+  }
+}
 
-    const RowVectorPtr writerInput = partitionSize == input->size()
-        ? input
-        : exec::wrap(partitionSize, partitionRows_[index], input);
-    write(index, writerInput);
+void IcebergDataSink::buildPartitionData(int32_t index) {
+  std::vector<folly::dynamic> partitionValues(partitionChannels_.size());
+  auto icebergPartitionIdGenerator =
+      dynamic_cast<const IcebergPartitionIdGenerator*>(
+          partitionIdGenerator_.get());
+  VELOX_CHECK_NOT_NULL(icebergPartitionIdGenerator);
+  const RowVectorPtr transformedValues =
+      icebergPartitionIdGenerator->partitionValues();
+  for (auto i = 0; i < partitionChannels_.size(); ++i) {
+    auto block = transformedValues->childAt(i);
+    if (block->isNullAt(index)) {
+      partitionValues[i] = nullptr;
+    } else {
+      DecodedVector decoded(*block);
+      partitionValues[i] = VELOX_DYNAMIC_SCALAR_TYPE_DISPATCH(
+          extractPartitionValue, block->typeKind(), &decoded, index);
+    }
   }
+  partitionData_[index] = partitionValues;
+}
+
+void IcebergDataSink::clusteredWrite(RowVectorPtr input, int32_t writerIdx) {
+  if (partitionSizes_[writerIdx] != 0) {
+    VELOX_CHECK_NOT_NULL(partitionRows_[writerIdx]);
+    partitionRows_[writerIdx]->setSize(
+        partitionSizes_[writerIdx] * sizeof(vector_size_t));
+  }
+  const vector_size_t partitionSize = partitionSizes_[writerIdx];
+  const RowVectorPtr writerInput = partitionSize == input->size()
+      ? input
+      : exec::wrap(partitionSize, partitionRows_[writerIdx], input);
+  write(writerIdx, writerInput);
 }
 
 HiveWriterId IcebergDataSink::getIcebergWriterId(size_t row) const {
@@ -470,9 +517,11 @@ void IcebergDataSink::closeInternal() {
 
   if (state_ == State::kClosed) {
     for (int i = 0; i < writers_.size(); ++i) {
-      WRITER_NON_RECLAIMABLE_SECTION_GUARD(i);
-      writers_[i]->close();
-      dataFileStats_.push_back(writers_[i]->dataFileStats());
+      if (writers_[i]) {
+        WRITER_NON_RECLAIMABLE_SECTION_GUARD(i);
+        writers_[i]->close();
+        dataFileStats_.push_back(writers_[i]->dataFileStats());
+      }
     }
   } else {
     for (int i = 0; i < writers_.size(); ++i) {
@@ -482,6 +531,63 @@ void IcebergDataSink::closeInternal() {
   }
 }
 
+void IcebergDataSink::closeWriter(int32_t index) {
+  common::testutil::TestValue::adjust(
+      "facebook::velox::connector::hive::iceberg::IcebergDataSink::closeWriter",
+      this);
+
+  if (writers_[index]) {
+    WRITER_NON_RECLAIMABLE_SECTION_GUARD(index);
+    if (sortWrite()) {
+      finishWriter(index);
+    }
+    writers_[index]->close();
+    dataFileStats_.push_back(writers_[index]->dataFileStats());
+    writers_[index] = nullptr;
+  }
+}
+
+bool IcebergDataSink::finishWriter(int32_t index) {
+  if (!sortWrite()) {
+    return true;
+  }
+
+  if (writers_[index]) {
+    const uint64_t startTimeMs = getCurrentTimeMs();
+    if (!writers_[index]->finish()) {
+      return false;
+    }
+    if (getCurrentTimeMs() - startTimeMs > sortWriterFinishTimeSliceLimitMs_) {
+      return false;
+    }
+  }
+  return true;
+}
+
+bool IcebergDataSink::finish() {
+  // Flush is reentry state.
+  setState(State::kFinishing);
+
+  // As for now, only sorted writer needs flush buffered data. For non-sorted
+  // writer, data is directly written to the underlying file writer.
+  if (!sortWrite()) {
+    return true;
+  }
+
+  // TODO: we might refactor to move the data sorting logic into hive data sink.
+  const uint64_t startTimeMs = getCurrentTimeMs();
+  for (auto i = 0; i < writers_.size(); ++i) {
+    WRITER_NON_RECLAIMABLE_SECTION_GUARD(i);
+    if (writers_[i] && !writers_[i]->finish()) {
+      return false;
+    }
+    if (getCurrentTimeMs() - startTimeMs > sortWriterFinishTimeSliceLimitMs_) {
+      return false;
+    }
+  }
+  return true;
+}
+
 std::unique_ptr<facebook::velox::dwio::common::Writer>
 IcebergDataSink::maybeCreateBucketSortWriter(
     std::unique_ptr<facebook::velox::dwio::common::Writer> writer) {

velox/connectors/hive/iceberg/IcebergDataSink.h

@@ -125,6 +125,8 @@ dataFileStats() const {
   /// Presto and Spark Iceberg commit protocol.
   std::vector<std::string> commitMessage() const override;
 
+  bool finish() override;
+
 private:
     IcebergDataSink(
         RowTypePtr inputType,
@@ -137,6 +139,8 @@ dataFileStats() const {
 
     void splitInputRowsAndEnsureWriters(RowVectorPtr input) override;
 
+  void computePartition(const RowVectorPtr& input);
+
   HiveWriterId getIcebergWriterId(size_t row) const;
 
   std::shared_ptr<dwio::common::WriterOptions> createWriterOptions()
@@ -148,8 +152,16 @@ dataFileStats() const {
   std::unique_ptr<dwio::common::Writer> maybeCreateBucketSortWriter(
       std::unique_ptr<dwio::common::Writer> writer) override;
 
+  void buildPartitionData(int32_t index);
+
+  void clusteredWrite(RowVectorPtr input, int32_t writerIdx);
+
   void closeInternal() override;
 
+  void closeWriter(int32_t index);
+
+  bool finishWriter(int32_t index);
+
   // Below are structures for partitions from all inputs. partitionData_
   // is indexed by partitionId.
   std::vector<std::vector<folly::dynamic>> partitionData_;
@@ -159,6 +171,11 @@ dataFileStats() const {
       std::vector<std::unique_ptr<dwio::common::DataFileStatsSettings>>>
       statsSettings_;
   std::unique_ptr<DataFileStatsCollector> icebergStatsCollector_;
+
+  // Below are structures for clustered mode writer.
+  const bool fanoutEnabled_;
+  uint32_t currentWriterId_;
+  std::unordered_set<uint32_t> completedWriterIds_;
 };
 
 } // namespace facebook::velox::connector::hive::iceberg

velox/connectors/hive/iceberg/IcebergPartitionIdGenerator.cpp

@@ -84,10 +84,10 @@ IcebergPartitionIdGenerator::IcebergPartitionIdGenerator(
     hashers_.emplace_back(
         exec::VectorHasher::create(columnTransform->resultType(), i++));
     VELOX_USER_CHECK(
-        exec::VectorHasher::typeKindSupportsValueIds(
-            columnTransform->resultType()->kind()),
-        "Unsupported partition type: {}.",
-        columnTransform->resultType()->toString());
+    hashers_.back()->typeSupportsValueIds(),
+    "Unsupported partition type: {}.",
+    columnTransform->resultType()->toString());
+
     partitionKeyTypes.emplace_back(columnTransform->resultType());
     std::string key =
         columnTransform->transformType() == TransformType::kIdentity

velox/connectors/hive/iceberg/tests/CMakeLists.txt

@@ -71,6 +71,7 @@ if(NOT VELOX_DISABLE_GOOGLETEST)
           IcebergTestBase.cpp
           IcebergTransformE2ETest.cpp
           IcebergTransformUnitTest.cpp
+          IcebergWriterModeTest.cpp
           Main.cpp
           Murmur3Test.cpp)
 

velox/connectors/hive/iceberg/tests/IcebergTestBase.h

@@ -63,9 +63,12 @@ class IcebergTestBase : public exec::test::HiveConnectorTestBase {
     const std::vector<PartitionField>& transformSpecs,
     const RowTypePtr& rowType);
 
+  void setupMemoryPools();
+
   dwio::common::FileFormat fileFormat_{dwio::common::FileFormat::PARQUET};
   RowTypePtr rowType_;
   std::shared_ptr<memory::MemoryPool> opPool_;
+  std::shared_ptr<config::ConfigBase> connectorSessionProperties_;
 
 private:
   IcebergInsertTableHandlePtr createIcebergInsertTableHandle(
@@ -77,10 +80,8 @@ class IcebergTestBase : public exec::test::HiveConnectorTestBase {
   std::vector<std::string> listPartitionDirectories(
       const std::string& dataPath);
 
-  void setupMemoryPools();
   std::shared_ptr<memory::MemoryPool> root_;
   std::shared_ptr<memory::MemoryPool> connectorPool_;
-  std::shared_ptr<config::ConfigBase> connectorSessionProperties_;
   std::shared_ptr<HiveConfig> connectorConfig_;
   std::unique_ptr<ConnectorQueryCtx> connectorQueryCtx_;
   VectorFuzzer::Options fuzzerOptions_;