[GLUTEN-4170][VL] Decouple partitions from plan to avoid driver stall…

…ed (apache#4177)

backends-clickhouse/src/main/scala/io/glutenproject/backendsapi/clickhouse/CHIteratorApi.scala

@@ -47,11 +47,6 @@ import scala.collection.JavaConverters._
 
 class CHIteratorApi extends IteratorApi with Logging with LogLevelUtil {
 
-  /**
-   * Generate native row partition.
-   *
-   * @return
-   */
   override def genSplitInfo(
       partition: InputPartition,
       partitionSchema: StructType,
@@ -95,6 +90,26 @@ class CHIteratorApi extends IteratorApi with Logging with LogLevelUtil {
     }
   }
 
+  /**
+   * Generate native row partition.
+   *
+   * @return
+   */
+  override def genPartitions(
+      wsCtx: WholeStageTransformContext,
+      splitInfos: Seq[Seq[SplitInfo]]): Seq[BaseGlutenPartition] = {
+    splitInfos.zipWithIndex.map {
+      case (splitInfos, index) =>
+        wsCtx.substraitContext.initSplitInfosIndex(0)
+        wsCtx.substraitContext.setSplitInfos(splitInfos)
+        val substraitPlan = wsCtx.root.toProtobuf
+        GlutenPartition(
+          index,
+          substraitPlan.toByteArray,
+          locations = splitInfos.flatMap(_.preferredLocations().asScala).toArray)
+    }
+  }
+
   /**
    * Generate Iterator[ColumnarBatch] for first stage.
    *
@@ -252,7 +267,7 @@ class CHIteratorApi extends IteratorApi with Logging with LogLevelUtil {
           GlutenPartition(
             index,
             substraitPlan.toByteArray,
-            splitInfo.preferredLocations().asScala.toArray)
+            locations = splitInfo.preferredLocations().asScala.toArray)
       }
     }(t => logInfo(s"Generating the Substrait plan took: $t ms."))
 

backends-clickhouse/src/test/scala/org/apache/spark/affinity/MixedAffinitySuite.scala

@@ -49,7 +49,7 @@ class MixedAffinitySuite extends QueryTest with SharedSparkSession {
     }
     val partition = GlutenMergeTreePartition(0, "", "", "", "fakePath", 0, 0)
     val locations = affinity.getNativeMergeTreePartitionLocations(partition)
-    val nativePartition = GlutenPartition(0, PlanBuilder.EMPTY_PLAN, locations)
+    val nativePartition = GlutenPartition(0, PlanBuilder.EMPTY_PLAN, locations = locations)
     assertResult(Set("forced_host_host-0")) {
       nativePartition.preferredLocations().toSet
     }

backends-velox/src/main/scala/io/glutenproject/backendsapi/velox/IteratorApiImpl.scala

@@ -21,7 +21,7 @@ import io.glutenproject.backendsapi.IteratorApi
 import io.glutenproject.execution._
 import io.glutenproject.metrics.IMetrics
 import io.glutenproject.substrait.plan.PlanNode
-import io.glutenproject.substrait.rel.{LocalFilesBuilder, SplitInfo}
+import io.glutenproject.substrait.rel.{LocalFilesBuilder, LocalFilesNode, SplitInfo}
 import io.glutenproject.substrait.rel.LocalFilesNode.ReadFileFormat
 import io.glutenproject.utils._
 import io.glutenproject.vectorized._
@@ -46,16 +46,12 @@ import java.lang.{Long => JLong}
 import java.nio.charset.StandardCharsets
 import java.time.ZoneOffset
 import java.util.{ArrayList => JArrayList, HashMap => JHashMap, Map => JMap}
+import java.util.concurrent.TimeUnit
 
 import scala.collection.JavaConverters._
 
 class IteratorApiImpl extends IteratorApi with Logging {
 
-  /**
-   * Generate native row partition.
-   *
-   * @return
-   */
   override def genSplitInfo(
       partition: InputPartition,
       partitionSchema: StructType,
@@ -79,6 +75,24 @@ class IteratorApiImpl extends IteratorApi with Logging {
     }
   }
 
+  /** Generate native row partition. */
+  override def genPartitions(
+      wsCtx: WholeStageTransformContext,
+      splitInfos: Seq[Seq[SplitInfo]]): Seq[BaseGlutenPartition] = {
+    // Only serialize plan once, save lots time when plan is complex.
+    val planByteArray = wsCtx.root.toProtobuf.toByteArray
+
+    splitInfos.zipWithIndex.map {
+      case (splitInfos, index) =>
+        GlutenPartition(
+          index,
+          planByteArray,
+          splitInfos.map(_.asInstanceOf[LocalFilesNode].toProtobuf.toByteArray).toArray,
+          splitInfos.flatMap(_.preferredLocations().asScala).toArray
+        )
+    }
+  }
+
   private def constructSplitInfo(schema: StructType, files: Array[PartitionedFile]) = {
     val paths = new JArrayList[String]()
     val starts = new JArrayList[JLong]
@@ -124,25 +138,34 @@ class IteratorApiImpl extends IteratorApi with Logging {
     transKernel.injectWriteFilesTempPath(path)
   }
 
-  /**
-   * Generate Iterator[ColumnarBatch] for first stage.
-   *
-   * @return
-   */
+  /** Generate Iterator[ColumnarBatch] for first stage. */
   override def genFirstStageIterator(
       inputPartition: BaseGlutenPartition,
       context: TaskContext,
       pipelineTime: SQLMetric,
       updateInputMetrics: (InputMetricsWrapper) => Unit,
       updateNativeMetrics: IMetrics => Unit,
       inputIterators: Seq[Iterator[ColumnarBatch]] = Seq()): Iterator[ColumnarBatch] = {
+    assert(
+      inputPartition.isInstanceOf[GlutenPartition],
+      "Velox backend only accept GlutenPartition.")
+
+    val beforeBuild = System.nanoTime()
     val columnarNativeIterators =
       new JArrayList[GeneralInIterator](inputIterators.map {
         iter => new ColumnarBatchInIterator(iter.asJava)
       }.asJava)
     val transKernel = NativePlanEvaluator.create()
+
+    val splitInfoByteArray = inputPartition
+      .asInstanceOf[GlutenPartition]
+      .splitInfosByteArray
     val resIter: GeneralOutIterator =
-      transKernel.createKernelWithBatchIterator(inputPartition.plan, columnarNativeIterators)
+      transKernel.createKernelWithBatchIterator(
+        inputPartition.plan,
+        splitInfoByteArray,
+        columnarNativeIterators)
+    pipelineTime += TimeUnit.NANOSECONDS.toMillis(System.nanoTime() - beforeBuild)
 
     Iterators
       .wrap(resIter.asScala)
@@ -160,11 +183,7 @@ class IteratorApiImpl extends IteratorApi with Logging {
 
   // scalastyle:off argcount
 
-  /**
-   * Generate Iterator[ColumnarBatch] for final stage.
-   *
-   * @return
-   */
+  /** Generate Iterator[ColumnarBatch] for final stage. */
   override def genFinalStageIterator(
       context: TaskContext,
       inputIterators: Seq[Iterator[ColumnarBatch]],
@@ -186,7 +205,10 @@ class IteratorApiImpl extends IteratorApi with Logging {
     val nativeResultIterator =
       transKernel.createKernelWithBatchIterator(
         rootNode.toProtobuf.toByteArray,
-        columnarNativeIterator)
+        // Final iterator does not contain scan split, so pass empty split info to native here.
+        new Array[Array[Byte]](0),
+        columnarNativeIterator
+      )
 
     Iterators
       .wrap(nativeResultIterator.asScala)

cpp/core/compute/Runtime.h

@@ -72,6 +72,8 @@ class Runtime : public std::enable_shared_from_this<Runtime> {
 
   virtual void injectWriteFilesTempPath(const std::string& path) = 0;
 
+  virtual void parseSplitInfo(const uint8_t* data, int32_t size) = 0;
+
   // Just for benchmark
   ::substrait::Plan& getPlan() {
     return substraitPlan_;
@@ -140,6 +142,7 @@ class Runtime : public std::enable_shared_from_this<Runtime> {
  protected:
   std::unique_ptr<ObjectStore> objStore_ = ObjectStore::create();
   ::substrait::Plan substraitPlan_;
+  std::vector<::substrait::ReadRel_LocalFiles> localFiles_;
   std::optional<std::string> writeFilesTempPath_;
   SparkTaskInfo taskInfo_;
   // Session conf map

cpp/core/jni/JniWrapper.cc

@@ -368,6 +368,7 @@ Java_io_glutenproject_vectorized_PlanEvaluatorJniWrapper_nativeCreateKernelWithI
     jobject wrapper,
     jlong memoryManagerHandle,
     jbyteArray planArr,
+    jobjectArray splitInfosArr,
     jobjectArray iterArr,
     jint stageId,
     jint partitionId,
@@ -381,10 +382,17 @@ Java_io_glutenproject_vectorized_PlanEvaluatorJniWrapper_nativeCreateKernelWithI
 
   auto spillDirStr = jStringToCString(env, spillDir);
 
+  for (jsize i = 0, splitInfoArraySize = env->GetArrayLength(splitInfosArr); i < splitInfoArraySize; i++) {
+    jbyteArray splitInfoArray = static_cast<jbyteArray>(env->GetObjectArrayElement(splitInfosArr, i));
+    jsize splitInfoSize = env->GetArrayLength(splitInfoArray);
+    auto splitInfoData = reinterpret_cast<const uint8_t*>(env->GetByteArrayElements(splitInfoArray, nullptr));
+    ctx->parseSplitInfo(splitInfoData, splitInfoSize);
+  }
+
   auto planData = reinterpret_cast<const uint8_t*>(env->GetByteArrayElements(planArr, nullptr));
   auto planSize = env->GetArrayLength(planArr);
-
   ctx->parsePlan(planData, planSize, {stageId, partitionId, taskId});
+
   auto& conf = ctx->getConfMap();
 
   // Handle the Java iters

cpp/velox/benchmarks/GenericBenchmark.cc

@@ -109,7 +109,8 @@ void populateWriterMetrics(
 } // namespace
 
 auto BM_Generic = [](::benchmark::State& state,
-                     const std::string& substraitJsonFile,
+                     const std::string& planFile,
+                     const std::string& splitFile,
                      const std::vector<std::string>& inputFiles,
                      const std::unordered_map<std::string, std::string>& conf,
                      FileReaderType readerType) {
@@ -119,11 +120,15 @@ auto BM_Generic = [](::benchmark::State& state,
   } else {
     setCpu(state.thread_index());
   }
+  bool emptySplit = splitFile.empty();
   memory::MemoryManager::testingSetInstance({});
   auto memoryManager = getDefaultMemoryManager();
   auto runtime = Runtime::create(kVeloxRuntimeKind, conf);
-  const auto& filePath = substraitJsonFile;
-  auto plan = getPlanFromFile(filePath);
+  auto plan = getPlanFromFile("Plan", planFile);
+  std::string split;
+  if (!emptySplit) {
+    split = getPlanFromFile("ReadRel.LocalFiles", splitFile);
+  }
   auto startTime = std::chrono::steady_clock::now();
   int64_t collectBatchTime = 0;
   WriterMetrics writerMetrics{};
@@ -145,6 +150,9 @@ auto BM_Generic = [](::benchmark::State& state,
     }
 
     runtime->parsePlan(reinterpret_cast<uint8_t*>(plan.data()), plan.size(), {});
+    if (!emptySplit) {
+      runtime->parseSplitInfo(reinterpret_cast<uint8_t*>(split.data()), split.size());
+    }
     auto resultIter =
         runtime->createResultIterator(memoryManager.get(), "/tmp/test-spill", std::move(inputIters), conf);
     auto veloxPlan = dynamic_cast<gluten::VeloxRuntime*>(runtime)->getVeloxPlan();
@@ -231,6 +239,7 @@ int main(int argc, char** argv) {
   gflags::ParseCommandLineFlags(&argc, &argv, true);
 
   std::string substraitJsonFile;
+  std::string splitFile;
   std::vector<std::string> inputFiles;
   std::unordered_map<std::string, std::string> conf;
 
@@ -242,18 +251,20 @@ int main(int argc, char** argv) {
     if (argc < 2) {
       LOG(INFO)
           << "No input args. Usage: " << std::endl
-          << "./generic_benchmark /absolute-path/to/substrait_json_file /absolute-path/to/data_file_1 /absolute-path/to/data_file_2 ...";
+          << "./generic_benchmark /absolute-path/to/substrait_json_file /absolute-path/to/split_json_file(optional)"
+          << " /absolute-path/to/data_file_1 /absolute-path/to/data_file_2 ...";
       LOG(INFO) << "Running example...";
       inputFiles.resize(2);
       substraitJsonFile = getGeneratedFilePath("example.json");
       inputFiles[0] = getGeneratedFilePath("example_orders");
       inputFiles[1] = getGeneratedFilePath("example_lineitem");
     } else {
       substraitJsonFile = argv[1];
+      splitFile = argv[2];
       abortIfFileNotExists(substraitJsonFile);
       LOG(INFO) << "Using substrait json file: " << std::endl << substraitJsonFile;
       LOG(INFO) << "Using " << argc - 2 << " input data file(s): ";
-      for (auto i = 2; i < argc; ++i) {
+      for (auto i = 3; i < argc; ++i) {
         inputFiles.emplace_back(argv[i]);
         abortIfFileNotExists(inputFiles.back());
         LOG(INFO) << inputFiles.back();
@@ -265,19 +276,20 @@ int main(int argc, char** argv) {
     std::exit(EXIT_FAILURE);
   }
 
-#define GENERIC_BENCHMARK(NAME, READER_TYPE)                                                                      \
-  do {                                                                                                            \
-    auto* bm = ::benchmark::RegisterBenchmark(NAME, BM_Generic, substraitJsonFile, inputFiles, conf, READER_TYPE) \
-                   ->MeasureProcessCPUTime()                                                                      \
-                   ->UseRealTime();                                                                               \
-    if (FLAGS_threads > 0) {                                                                                      \
-      bm->Threads(FLAGS_threads);                                                                                 \
-    } else {                                                                                                      \
-      bm->ThreadRange(1, std::thread::hardware_concurrency());                                                    \
-    }                                                                                                             \
-    if (FLAGS_iterations > 0) {                                                                                   \
-      bm->Iterations(FLAGS_iterations);                                                                           \
-    }                                                                                                             \
+#define GENERIC_BENCHMARK(NAME, READER_TYPE)                                                                          \
+  do {                                                                                                                \
+    auto* bm =                                                                                                        \
+        ::benchmark::RegisterBenchmark(NAME, BM_Generic, substraitJsonFile, splitFile, inputFiles, conf, READER_TYPE) \
+            ->MeasureProcessCPUTime()                                                                                 \
+            ->UseRealTime();                                                                                          \
+    if (FLAGS_threads > 0) {                                                                                          \
+      bm->Threads(FLAGS_threads);                                                                                     \
+    } else {                                                                                                          \
+      bm->ThreadRange(1, std::thread::hardware_concurrency());                                                        \
+    }                                                                                                                 \
+    if (FLAGS_iterations > 0) {                                                                                       \
+      bm->Iterations(FLAGS_iterations);                                                                               \
+    }                                                                                                                 \
   } while (0)
 
   DLOG(INFO) << "FLAGS_threads:" << FLAGS_threads;

cpp/velox/benchmarks/common/BenchmarkUtils.cc

@@ -47,14 +47,14 @@ void initVeloxBackend() {
   initVeloxBackend(bmConfMap);
 }
 
-std::string getPlanFromFile(const std::string& filePath) {
+std::string getPlanFromFile(const std::string& type, const std::string& filePath) {
   // Read json file and resume the binary data.
   std::ifstream msgJson(filePath);
   std::stringstream buffer;
   buffer << msgJson.rdbuf();
   std::string msgData = buffer.str();
 
-  return gluten::substraitFromJsonToPb("Plan", msgData);
+  return gluten::substraitFromJsonToPb(type, msgData);
 }
 
 velox::dwio::common::FileFormat getFileFormat(const std::string& fileFormat) {

cpp/velox/benchmarks/common/BenchmarkUtils.h

@@ -72,7 +72,7 @@ inline std::string getGeneratedFilePath(const std::string& fileName) {
 }
 
 /// Read binary data from a json file.
-std::string getPlanFromFile(const std::string& filePath);
+std::string getPlanFromFile(const std::string& type, const std::string& filePath);
 
 /// Get the file paths, starts, lengths from a directory.
 /// Use fileFormat to specify the format to read, eg., orc, parquet.