[SPARK-22908] Add kafka source and sink for continuous processing.

external/kafka-0-10-sql/src/main/scala/org/apache/spark/sql/kafka010/KafkaContinuousReader.scala

@@ -0,0 +1,232 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.spark.sql.kafka010
+
+import java.{util => ju}
+
+import org.apache.kafka.clients.consumer.ConsumerRecord
+import org.apache.kafka.common.TopicPartition
+
+import org.apache.spark.internal.Logging
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.catalyst.expressions.UnsafeRow
+import org.apache.spark.sql.catalyst.expressions.codegen.{BufferHolder, UnsafeRowWriter}
+import org.apache.spark.sql.catalyst.util.DateTimeUtils
+import org.apache.spark.sql.kafka010.KafkaSource.{INSTRUCTION_FOR_FAIL_ON_DATA_LOSS_FALSE, INSTRUCTION_FOR_FAIL_ON_DATA_LOSS_TRUE}
+import org.apache.spark.sql.sources.v2.reader._
+import org.apache.spark.sql.sources.v2.streaming.reader.{ContinuousDataReader, ContinuousReader, Offset, PartitionOffset}
+import org.apache.spark.sql.types.StructType
+import org.apache.spark.unsafe.types.UTF8String
+
+/**
+ * A [[ContinuousReader]] for data from kafka.
+ *
+ * @param offsetReader  a reader used to get kafka offsets. Note that the actual data will be
+ *                      read by per-task consumers generated later.
+ * @param kafkaParams   String params for per-task Kafka consumers.
+ * @param sourceOptions The [[org.apache.spark.sql.sources.v2.DataSourceV2Options]] params which
+ *                      are not Kafka consumer params.
+ * @param metadataPath Path to a directory this reader can use for writing metadata.
+ * @param initialOffsets The Kafka offsets to start reading data at.
+ * @param failOnDataLoss Flag indicating whether reading should fail in data loss
+ *                       scenarios, where some offsets after the specified initial ones can't be
+ *                       properly read.
+ */
+class KafkaContinuousReader(
+    offsetReader: KafkaOffsetReader,
+    kafkaParams: ju.Map[String, Object],
+    sourceOptions: Map[String, String],
+    metadataPath: String,
+    initialOffsets: KafkaOffsetRangeLimit,
+    failOnDataLoss: Boolean)
+  extends ContinuousReader with SupportsScanUnsafeRow with Logging {
+
+  private lazy val session = SparkSession.getActiveSession.get
+  private lazy val sc = session.sparkContext
+
+  // Initialized when creating read tasks. If this diverges from the partitions at the latest
+  // offsets, we need to reconfigure.
+  // Exposed outside this object only for unit tests.
+  private[sql] var knownPartitions: Set[TopicPartition] = _
+
+  override def readSchema: StructType = KafkaOffsetReader.kafkaSchema
+
+  private var offset: Offset = _
+  override def setOffset(start: ju.Optional[Offset]): Unit = {
+    offset = start.orElse {
+      val offsets = initialOffsets match {
+        case EarliestOffsetRangeLimit => KafkaSourceOffset(offsetReader.fetchEarliestOffsets())
+        case LatestOffsetRangeLimit => KafkaSourceOffset(offsetReader.fetchLatestOffsets())
+        case SpecificOffsetRangeLimit(p) => offsetReader.fetchSpecificOffsets(p, reportDataLoss)
+      }
+      logInfo(s"Initial offsets: $offsets")
+      offsets
+    }
+  }
+
+  override def getStartOffset(): Offset = offset
+
+  override def deserializeOffset(json: String): Offset = {
+    KafkaSourceOffset(JsonUtils.partitionOffsets(json))
+  }
+
+  override def createUnsafeRowReadTasks(): ju.List[ReadTask[UnsafeRow]] = {
+    import scala.collection.JavaConverters._
+
+    val oldStartPartitionOffsets = KafkaSourceOffset.getPartitionOffsets(offset)
+
+    val currentPartitionSet = offsetReader.fetchEarliestOffsets().keySet
+    val newPartitions = currentPartitionSet.diff(oldStartPartitionOffsets.keySet)
+    val newPartitionOffsets = offsetReader.fetchEarliestOffsets(newPartitions.toSeq)
+
+    val deletedPartitions = oldStartPartitionOffsets.keySet.diff(currentPartitionSet)
+    if (deletedPartitions.nonEmpty) {
+      reportDataLoss(s"Some partitions were deleted: $deletedPartitions")
+    }
+
+    val startOffsets = newPartitionOffsets ++
+      oldStartPartitionOffsets.filterKeys(!deletedPartitions.contains(_))
+    knownPartitions = startOffsets.keySet
+
+    startOffsets.toSeq.map {
+      case (topicPartition, start) =>
+        KafkaContinuousReadTask(
+          topicPartition, start, kafkaParams, failOnDataLoss)
+          .asInstanceOf[ReadTask[UnsafeRow]]
+    }.asJava
+  }
+
+  /** Stop this source and free any resources it has allocated. */
+  def stop(): Unit = synchronized {
+    offsetReader.close()
+  }
+
+  override def commit(end: Offset): Unit = {}
+
+  override def mergeOffsets(offsets: Array[PartitionOffset]): Offset = {
+    val mergedMap = offsets.map {
+      case KafkaSourcePartitionOffset(p, o) => Map(p -> o)
+    }.reduce(_ ++ _)
+    KafkaSourceOffset(mergedMap)
+  }
+
+  override def needsReconfiguration(): Boolean = {
+    knownPartitions != null && offsetReader.fetchLatestOffsets().keySet != knownPartitions
+  }
+
+  override def toString(): String = s"KafkaSource[$offsetReader]"
+
+  /**
+   * If `failOnDataLoss` is true, this method will throw an `IllegalStateException`.
+   * Otherwise, just log a warning.
+   */
+  private def reportDataLoss(message: String): Unit = {
+    if (failOnDataLoss) {
+      throw new IllegalStateException(message + s". $INSTRUCTION_FOR_FAIL_ON_DATA_LOSS_TRUE")
+    } else {
+      logWarning(message + s". $INSTRUCTION_FOR_FAIL_ON_DATA_LOSS_FALSE")
+    }
+  }
+}
+
+/**
+ * A read task for continuous Kafka processing. This will be serialized and transformed into a
+ * full reader on executors.
+ *
+ * @param topicPartition The (topic, partition) pair this task is responsible for.
+ * @param startOffset The offset to start reading from within the partition.
+ * @param kafkaParams Kafka consumer params to use.
+ * @param failOnDataLoss Flag indicating whether data reader should fail if some offsets
+ *                       are skipped.
+ */
+case class KafkaContinuousReadTask(
+    topicPartition: TopicPartition,
+    startOffset: Long,
+    kafkaParams: ju.Map[String, Object],
+    failOnDataLoss: Boolean) extends ReadTask[UnsafeRow] {
+  override def createDataReader(): KafkaContinuousDataReader = {
+    new KafkaContinuousDataReader(topicPartition, startOffset, kafkaParams, failOnDataLoss)
+  }
+}
+
+/**
+ * A per-task data reader for continuous Kafka processing.
+ *
+ * @param topicPartition The (topic, partition) pair this data reader is responsible for.
+ * @param startOffset The offset to start reading from within the partition.
+ * @param kafkaParams Kafka consumer params to use.
+ * @param failOnDataLoss Flag indicating whether data reader should fail if some offsets
+ *                       are skipped.
+ */
+class KafkaContinuousDataReader(
+    topicPartition: TopicPartition,
+    startOffset: Long,
+    kafkaParams: ju.Map[String, Object],
+    failOnDataLoss: Boolean) extends ContinuousDataReader[UnsafeRow] {
+  private val topic = topicPartition.topic
+  private val kafkaPartition = topicPartition.partition
+  private val consumer = CachedKafkaConsumer.createUncached(topic, kafkaPartition, kafkaParams)
+
+  private val sharedRow = new UnsafeRow(7)
+  private val bufferHolder = new BufferHolder(sharedRow)
+  private val rowWriter = new UnsafeRowWriter(bufferHolder, 7)
+
+  private var nextKafkaOffset = startOffset
+  private var currentRecord: ConsumerRecord[Array[Byte], Array[Byte]] = _
+
+  override def next(): Boolean = {
+    var r: ConsumerRecord[Array[Byte], Array[Byte]] = null
+    while (r == null) {
+      r = consumer.get(
+        nextKafkaOffset,
+        untilOffset = Long.MaxValue,
+        pollTimeoutMs = Long.MaxValue,
+        failOnDataLoss)
+    }
+    nextKafkaOffset = r.offset + 1
+    currentRecord = r
+    true
+  }
+
+  override def get(): UnsafeRow = {
+    bufferHolder.reset()
+
+    if (currentRecord.key == null) {
+      rowWriter.setNullAt(0)
+    } else {
+      rowWriter.write(0, currentRecord.key)
+    }
+    rowWriter.write(1, currentRecord.value)
+    rowWriter.write(2, UTF8String.fromString(currentRecord.topic))
+    rowWriter.write(3, currentRecord.partition)
+    rowWriter.write(4, currentRecord.offset)
+    rowWriter.write(5,
+      DateTimeUtils.fromJavaTimestamp(new java.sql.Timestamp(currentRecord.timestamp)))
+    rowWriter.write(6, currentRecord.timestampType.id)
+    sharedRow.setTotalSize(bufferHolder.totalSize)
+    sharedRow
+  }
+
+  override def getOffset(): KafkaSourcePartitionOffset = {
+    KafkaSourcePartitionOffset(topicPartition, nextKafkaOffset)
+  }
+
+  override def close(): Unit = {
+    consumer.close()
+  }
+}

external/kafka-0-10-sql/src/main/scala/org/apache/spark/sql/kafka010/KafkaContinuousWriter.scala

@@ -0,0 +1,119 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.spark.sql.kafka010
+
+import org.apache.kafka.clients.producer.{Callback, ProducerRecord, RecordMetadata}
+import scala.collection.JavaConverters._
+
+import org.apache.spark.internal.Logging
+import org.apache.spark.sql.{Row, SparkSession}
+import org.apache.spark.sql.catalyst.InternalRow
+import org.apache.spark.sql.catalyst.expressions.{Attribute, Cast, Literal, UnsafeProjection}
+import org.apache.spark.sql.kafka010.KafkaSourceProvider.{kafkaParamsForProducer, TOPIC_OPTION_KEY}
+import org.apache.spark.sql.kafka010.KafkaWriter.validateQuery
+import org.apache.spark.sql.sources.v2.streaming.writer.ContinuousWriter
+import org.apache.spark.sql.sources.v2.writer._
+import org.apache.spark.sql.streaming.OutputMode
+import org.apache.spark.sql.types.{BinaryType, StringType, StructType}
+
+/**
+ * Dummy commit message. The DataSourceV2 framework requires a commit message implementation but we
+ * don't need to really send one.
+ */
+case object KafkaWriterCommitMessage extends WriterCommitMessage
+
+/**
+ * A [[ContinuousWriter]] for Kafka writing. Responsible for generating the writer factory.
+ * @param topic The topic this writer is responsible for. If None, topic will be inferred from
+ *              a `topic` field in the incoming data.
+ * @param producerParams Parameters for Kafka producers in each task.
+ * @param schema The schema of the input data.
+ */
+class KafkaContinuousWriter(
+    topic: Option[String], producerParams: Map[String, String], schema: StructType)
+  extends ContinuousWriter with SupportsWriteInternalRow {
+
+  validateQuery(schema.toAttributes, producerParams.toMap[String, Object].asJava, topic)
+
+  override def createInternalRowWriterFactory(): KafkaContinuousWriterFactory =
+    KafkaContinuousWriterFactory(topic, producerParams, schema)
+
+  override def commit(epochId: Long, messages: Array[WriterCommitMessage]): Unit = {}
+  override def abort(messages: Array[WriterCommitMessage]): Unit = {}
+}
+
+/**
+ * A [[DataWriterFactory]] for Kafka writing. Will be serialized and sent to executors to generate
+ * the per-task data writers.
+ * @param topic The topic that should be written to. If None, topic will be inferred from
+ *              a `topic` field in the incoming data.
+ * @param producerParams Parameters for Kafka producers in each task.
+ * @param schema The schema of the input data.
+ */
+case class KafkaContinuousWriterFactory(
+    topic: Option[String], producerParams: Map[String, String], schema: StructType)
+  extends DataWriterFactory[InternalRow] {
+
+  override def createDataWriter(partitionId: Int, attemptNumber: Int): DataWriter[InternalRow] = {
+    new KafkaContinuousDataWriter(topic, producerParams, schema.toAttributes)
+  }
+}
+
+/**
+ * A [[DataWriter]] for Kafka writing. One data writer will be created in each partition to
+ * process incoming rows.
+ *
+ * @param targetTopic The topic that this data writer is targeting. If None, topic will be inferred
+ *                    from a `topic` field in the incoming data.
+ * @param producerParams Parameters to use for the Kafka producer.
+ * @param inputSchema The attributes in the input data.
+ */
+class KafkaContinuousDataWriter(
+    targetTopic: Option[String], producerParams: Map[String, String], inputSchema: Seq[Attribute])
+  extends KafkaRowWriter(inputSchema, targetTopic) with DataWriter[InternalRow] {
+  import scala.collection.JavaConverters._
+
+  private lazy val producer = CachedKafkaProducer.getOrCreate(
+    new java.util.HashMap[String, Object](producerParams.asJava))
+
+  def write(row: InternalRow): Unit = {
+    checkForErrors()
+    sendRow(row, producer)
+  }
+
+  def commit(): WriterCommitMessage = {
+    // Send is asynchronous, but we can't commit until all rows are actually in Kafka.
+    // This requires flushing and then checking that no callbacks produced errors.
+    // We also check for errors before to fail as soon as possible - the check is cheap.
+    checkForErrors()
+    producer.flush()
+    checkForErrors()
+    KafkaWriterCommitMessage
+  }
+
+  def abort(): Unit = {}
+
+  def close(): Unit = {
+    checkForErrors()
+    if (producer != null) {
+      producer.flush()
+      checkForErrors()
+      CachedKafkaProducer.close(new java.util.HashMap[String, Object](producerParams.asJava))
+    }
+  }
+}

external/kafka-0-10-sql/src/main/scala/org/apache/spark/sql/kafka010/KafkaOffsetReader.scala

@@ -117,10 +117,14 @@ private[kafka010] class KafkaOffsetReader(
    * Resolves the specific offsets based on Kafka seek positions.
    * This method resolves offset value -1 to the latest and -2 to the
    * earliest Kafka seek position.
+   *
+   * @param partitionOffsets the specific offsets to resolve
+   * @param reportDataLoss callback to either report or log data loss depending on setting
    */
   def fetchSpecificOffsets(
-      partitionOffsets: Map[TopicPartition, Long]): Map[TopicPartition, Long] =
-    runUninterruptibly {
+      partitionOffsets: Map[TopicPartition, Long],
+      reportDataLoss: String => Unit): KafkaSourceOffset = {
+    val fetched = runUninterruptibly {
       withRetriesWithoutInterrupt {
         // Poll to get the latest assigned partitions
         consumer.poll(0)
@@ -145,6 +149,19 @@ private[kafka010] class KafkaOffsetReader(
       }
     }
 
+    partitionOffsets.foreach {
+      case (tp, off) if off != KafkaOffsetRangeLimit.LATEST &&
+        off != KafkaOffsetRangeLimit.EARLIEST =>
+        if (fetched(tp) != off) {
+          reportDataLoss(
+            s"startingOffsets for $tp was $off but consumer reset to ${fetched(tp)}")
+        }
+      case _ =>
+        // no real way to check that beginning or end is reasonable
+    }
+    KafkaSourceOffset(fetched)
+  }
+
   /**
    * Fetch the earliest offsets for the topic partitions that are indicated
    * in the [[ConsumerStrategy]].