arrow zerocopy for read and write in object store

Author: Deegue

core/raydp-main/src/main/java/org/apache/spark/raydp/RayExecutorUtils.java

@@ -21,11 +21,12 @@
 import io.ray.api.ObjectRef;
 import io.ray.api.Ray;
 import io.ray.api.call.ActorCreator;
+import io.ray.api.placementgroup.PlacementGroup;
+import io.ray.runtime.object.ObjectRefImpl;
+
 import java.util.Map;
 import java.util.List;
 
-import io.ray.api.placementgroup.PlacementGroup;
-import io.ray.runtime.object.ObjectRefImpl;
 import org.apache.spark.executor.RayDPExecutor;
 
 public class RayExecutorUtils {

core/raydp-main/src/main/scala/org/apache/spark/rdd/RayDatasetRDD.scala

@@ -22,6 +22,7 @@ import java.util.List;
 import scala.collection.JavaConverters._
 
 import io.ray.runtime.generated.Common.Address
+import org.apache.arrow.vector.VectorSchemaRoot
 
 import org.apache.spark.{Partition, SparkContext, TaskContext}
 import org.apache.spark.api.java.JavaSparkContext
@@ -37,15 +38,15 @@ class RayDatasetRDD(
     jsc: JavaSparkContext,
     @transient val objectIds: List[Array[Byte]],
     locations: List[Array[Byte]])
-  extends RDD[Array[Byte]](jsc.sc, Nil) {
+  extends RDD[VectorSchemaRoot](jsc.sc, Nil) {
 
   override def getPartitions: Array[Partition] = {
     objectIds.asScala.zipWithIndex.map { case (k, i) =>
       new RayDatasetRDDPartition(k, i).asInstanceOf[Partition]
     }.toArray
   }
 
-  override def compute(split: Partition, context: TaskContext): Iterator[Array[Byte]] = {
+  override def compute(split: Partition, context: TaskContext): Iterator[VectorSchemaRoot] = {
     val ref = split.asInstanceOf[RayDatasetRDDPartition].ref
     ObjectStoreReader.getBatchesFromStream(ref)
   }

core/raydp-main/src/main/scala/org/apache/spark/sql/raydp/ObjectStoreReader.scala

@@ -18,18 +18,26 @@
 package org.apache.spark.sql.raydp
 
 import java.io.ByteArrayInputStream
+import java.nio.ByteBuffer
 import java.nio.channels.{Channels, ReadableByteChannel}
 import java.util.List
 
+import scala.collection.JavaConverters._
+
 import com.intel.raydp.shims.SparkShimLoader
+import org.apache.arrow.vector.VectorSchemaRoot
 
-import org.apache.spark.api.java.{JavaRDD, JavaSparkContext}
+import org.apache.spark.TaskContext
+import org.apache.spark.api.java.JavaRDD
 import org.apache.spark.raydp.RayDPUtils
 import org.apache.spark.rdd.{RayDatasetRDD, RayObjectRefRDD}
 import org.apache.spark.sql.{DataFrame, SparkSession, SQLContext}
+import org.apache.spark.sql.catalyst.InternalRow
 import org.apache.spark.sql.catalyst.expressions.GenericRow
 import org.apache.spark.sql.execution.arrow.ArrowConverters
-import org.apache.spark.sql.types.{IntegerType, StructType}
+import org.apache.spark.sql.types._
+import org.apache.spark.sql.util.ArrowUtils
+import org.apache.spark.sql.vectorized.{ArrowColumnVector, ColumnarBatch, ColumnVector}
 
 object ObjectStoreReader {
   def createRayObjectRefDF(
@@ -40,17 +48,56 @@ object ObjectStoreReader {
     spark.createDataFrame(rdd, schema)
   }
 
+  def fromRootIterator(
+      arrowRootIter: Iterator[VectorSchemaRoot],
+      schema: StructType,
+      timeZoneId: String): Iterator[InternalRow] = {
+    val arrowSchema = ArrowUtils.toArrowSchema(schema, timeZoneId)
+
+    new Iterator[InternalRow] {
+      private var rowIter = if (arrowRootIter.hasNext) nextBatch() else Iterator.empty
+
+      override def hasNext: Boolean = rowIter.hasNext || {
+        if (arrowRootIter.hasNext) {
+          rowIter = nextBatch()
+          true
+        } else {
+          false
+        }
+      }
+
+      override def next(): InternalRow = rowIter.next()
+
+      private def nextBatch(): Iterator[InternalRow] = {
+        val root = arrowRootIter.next()
+        val columns = root.getFieldVectors.asScala.map { vector =>
+          new ArrowColumnVector(vector).asInstanceOf[ColumnVector]
+        }.toArray
+
+        val batch = new ColumnarBatch(columns)
+        batch.setNumRows(root.getRowCount)
+        root.close()
+        batch.rowIterator().asScala
+      }
+    }
+  }
+
   def RayDatasetToDataFrame(
       sparkSession: SparkSession,
       rdd: RayDatasetRDD,
-      schema: String): DataFrame = {
-    SparkShimLoader.getSparkShims.toDataFrame(JavaRDD.fromRDD(rdd), schema, sparkSession)
+      schemaString: String): DataFrame = {
+    val schema = DataType.fromJson(schemaString).asInstanceOf[StructType]
+    val sqlContext = new SQLContext(sparkSession)
+    val timeZoneId = sqlContext.sessionState.conf.sessionLocalTimeZone
+    val resultRDD = JavaRDD.fromRDD(rdd).rdd.mapPartitions { it =>
+      fromRootIterator(it, schema, timeZoneId)
+    }
+    sqlContext.internalCreateDataFrame(resultRDD.setName("arrow"), schema)
   }
 
   def getBatchesFromStream(
-      ref: Array[Byte]): Iterator[Array[Byte]] = {
-    val objectRef = RayDPUtils.readBinary(ref, classOf[Array[Byte]])
-    ArrowConverters.getBatchesFromStream(
-        Channels.newChannel(new ByteArrayInputStream(objectRef.get)))
+      ref: Array[Byte]): Iterator[VectorSchemaRoot] = {
+    val objectRef = RayDPUtils.readBinary(ref, classOf[VectorSchemaRoot])
+    Iterator[VectorSchemaRoot](objectRef.get)
   }
 }

core/raydp-main/src/main/scala/org/apache/spark/sql/raydp/ObjectStoreWriter.scala

@@ -17,7 +17,6 @@
 
 package org.apache.spark.sql.raydp
 
-
 import java.io.ByteArrayOutputStream
 import java.util.{List, UUID}
 import java.util.concurrent.{ConcurrentHashMap, ConcurrentLinkedQueue}
@@ -61,17 +60,16 @@ class ObjectStoreWriter(@transient val df: DataFrame) extends Serializable {
   val uuid: UUID = ObjectStoreWriter.dfToId.getOrElseUpdate(df, UUID.randomUUID())
 
   def writeToRay(
-      data: Array[Byte],
+      root: VectorSchemaRoot,
       numRecords: Int,
       queue: ObjectRefHolder.Queue,
       ownerName: String): RecordBatch = {
-
-    var objectRef: ObjectRef[Array[Byte]] = null
+    var objectRef: ObjectRef[VectorSchemaRoot] = null
     if (ownerName == "") {
-      objectRef = Ray.put(data)
+      objectRef = Ray.put(root)
     } else {
       var dataOwner: PyActorHandle = Ray.getActor(ownerName).get()
-      objectRef = Ray.put(data, dataOwner)
+      objectRef = Ray.put(root, dataOwner)
     }
 
     // add the objectRef to the objectRefHolder to avoid reference GC
@@ -111,21 +109,15 @@ class ObjectStoreWriter(@transient val df: DataFrame) extends Serializable {
       val root = VectorSchemaRoot.create(arrowSchema, allocator)
       val results = new ArrayBuffer[RecordBatch]()
 
-      val byteOut = new ByteArrayOutputStream()
       val arrowWriter = ArrowWriter.create(root)
       var numRecords: Int = 0
 
       Utils.tryWithSafeFinally {
         while (batchIter.hasNext) {
           // reset the state
           numRecords = 0
-          byteOut.reset()
           arrowWriter.reset()
 
-          // write out the schema meta data
-          val writer = new ArrowStreamWriter(root, null, byteOut)
-          writer.start()
-
           // get the next record batch
           val nextBatch = batchIter.next()
 
@@ -136,19 +128,11 @@ class ObjectStoreWriter(@transient val df: DataFrame) extends Serializable {
 
           // set the write record count
           arrowWriter.finish()
-          // write out the record batch to the underlying out
-          writer.writeBatch()
-
-          // get the wrote ByteArray and save to Ray ObjectStore
-          val byteArray = byteOut.toByteArray
-          results += writeToRay(byteArray, numRecords, queue, ownerName)
-          // end writes footer to the output stream and doesn't clean any resources.
-          // It could throw exception if the output stream is closed, so it should be
-          // in the try block.
-          writer.end()
+
+          // write and schema root directly and save to Ray ObjectStore
+          results += writeToRay(root, numRecords, queue, ownerName)
         }
         arrowWriter.reset()
-        byteOut.close()
       } {
         // If we close root and allocator in TaskCompletionListener, there could be a race
         // condition where the writer thread keeps writing to the VectorSchemaRoot while
@@ -173,7 +157,7 @@ class ObjectStoreWriter(@transient val df: DataFrame) extends Serializable {
   /**
    * For test.
    */
-  def getRandomRef(): List[Array[Byte]] = {
+  def getRandomRef(): List[VectorSchemaRoot] = {
 
     df.queryExecution.toRdd.mapPartitions { _ =>
       Iterator(ObjectRefHolder.getRandom(uuid))
@@ -233,7 +217,7 @@ object ObjectStoreWriter {
     var executorIds = df.sqlContext.sparkContext.getExecutorIds.toArray
     val numExecutors = executorIds.length
     val appMasterHandle = Ray.getActor(RayAppMaster.ACTOR_NAME)
-                             .get.asInstanceOf[ActorHandle[RayAppMaster]]
+                            .get.asInstanceOf[ActorHandle[RayAppMaster]]
     val restartedExecutors = RayAppMasterUtils.getRestartedExecutors(appMasterHandle)
     // Check if there is any restarted executors
     if (!restartedExecutors.isEmpty) {
@@ -251,8 +235,8 @@ object ObjectStoreWriter {
     val refs = new Array[ObjectRef[Array[Byte]]](numPartitions)
     val handles = executorIds.map {id =>
       Ray.getActor("raydp-executor-" + id)
-         .get
-         .asInstanceOf[ActorHandle[RayDPExecutor]]
+        .get
+        .asInstanceOf[ActorHandle[RayDPExecutor]]
     }
     val handlesMap = (executorIds zip handles).toMap
     val locations = RayExecutorUtils.getBlockLocations(
@@ -261,18 +245,15 @@ object ObjectStoreWriter {
       // TODO use getPreferredLocs, but we don't have a host ip to actor table now
       refs(i) = RayExecutorUtils.getRDDPartition(
           handlesMap(locations(i)), rdd.id, i, schema, driverAgentUrl)
-      queue.add(refs(i))
-    }
-    for (i <- 0 until numPartitions) {
+      queue.add(RayDPUtils.readBinary(refs(i).get(), classOf[VectorSchemaRoot]))
       results(i) = RayDPUtils.convert(refs(i)).getId.getBytes
     }
     results
   }
-
 }
 
 object ObjectRefHolder {
-  type Queue = ConcurrentLinkedQueue[ObjectRef[Array[Byte]]]
+  type Queue = ConcurrentLinkedQueue[ObjectRef[VectorSchemaRoot]]
   private val dfToQueue = new ConcurrentHashMap[UUID, Queue]()
 
   def getQueue(df: UUID): Queue = {
@@ -297,7 +278,7 @@ object ObjectRefHolder {
     queue.size()
   }
 
-  def getRandom(df: UUID): Array[Byte] = {
+  def getRandom(df: UUID): VectorSchemaRoot = {
     val queue = checkQueueExists(df)
     val ref = RayDPUtils.convert(queue.peek())
     ref.get()

python/raydp/spark/dataset.py

@@ -237,7 +237,7 @@ def _convert_blocks_to_dataframe(blocks):
     return df
 
 def _convert_by_rdd(spark: sql.SparkSession,
-                    blocks: Dataset,
+                    blocks: List[ObjectRef],
                     locations: List[bytes],
                     schema: StructType) -> DataFrame:
     object_ids = [block.binary() for block in blocks]
@@ -269,14 +269,7 @@ def ray_dataset_to_spark_dataframe(spark: sql.SparkSession,
     schema = StructType()
     for field in arrow_schema:
         schema.add(field.name, from_arrow_type(field.type), nullable=field.nullable)
-    #TODO how to branch on type of block?
-    sample = ray.get(blocks[0])
-    if isinstance(sample, bytes):
-        return _convert_by_rdd(spark, blocks, locations, schema)
-    elif isinstance(sample, pa.Table):
-        return _convert_by_udf(spark, blocks, locations, schema)
-    else:
-        raise RuntimeError("ray.to_spark only supports arrow type blocks")
+    return _convert_by_rdd(spark, blocks, locations, schema)
 
 if HAS_MLDATASET:
     class RecordBatch(_SourceShard):