Github user tdas commented on a diff in the pull request:
https://github.com/apache/spark/pull/807#discussion_r13838881
--- Diff:
external/flume-sink/src/main/scala/org/apache/spark/flume/sink/SparkSink.scala
---
@@ -0,0 +1,432 @@
+/*
+ * 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.flume.sink
+
+import java.net.InetSocketAddress
+import java.nio.ByteBuffer
+import java.util
+import java.util.concurrent._
+import java.util.concurrent.atomic.AtomicLong
+
+import scala.util.control.Breaks
+
+import com.google.common.util.concurrent.ThreadFactoryBuilder
+import org.apache.avro.ipc.NettyServer
+import org.apache.avro.ipc.specific.SpecificResponder
+import org.apache.commons.lang.RandomStringUtils
+import org.apache.flume.Sink.Status
+import org.apache.flume.conf.{ConfigurationException, Configurable}
+import org.apache.flume.sink.AbstractSink
+import org.apache.flume.{Channel, Transaction, FlumeException, Context}
+import org.slf4j.LoggerFactory
+
+import org.apache.spark.flume.{SparkSinkEvent, EventBatch,
SparkFlumeProtocol}
+
+/**
+ * A sink that uses Avro RPC to run a server that can be polled by Spark's
+ * FlumePollingInputDStream. This sink has the following configuration
parameters:
+ *
+ * hostname - The hostname to bind to. Default: 0.0.0.0
+ * port - The port to bind to. (No default - mandatory)
+ * timeout - Time in seconds after which a transaction is rolled back,
+ * if an ACK is not received from Spark within that time
+ * threads - Number of threads to use to receive requests from Spark
(Default: 10)
+ *
+ */
+// Flume forces transactions to be thread-local. So each transaction
*must* be committed, or
+// rolled back from the thread it was originally created in. So each
getEvents call from Spark
+// creates a TransactionProcessor which runs in a new thread, in which the
transaction is created
+// and events are pulled off the channel. Once the events are sent to
spark,
+// that thread is blocked and the TransactionProcessor is saved in a map,
+// until an ACK or NACK comes back or the transaction times out (after the
specified timeout).
+// When the response comes, the TransactionProcessor is retrieved and then
unblocked,
+// at which point the transaction is committed or rolled back.
+class SparkSink extends AbstractSink with Configurable {
+
+ // Size of the pool to use for holding transaction processors.
+ private var poolSize: Integer = SparkSinkConfig.DEFAULT_THREADS
+
+ // Timeout for each transaction. If spark does not respond in this much
time,
+ // rollback the transaction
+ private var transactionTimeout =
SparkSinkConfig.DEFAULT_TRANSACTION_TIMEOUT
+
+ // Address info to bind on
+ private var hostname: String = SparkSinkConfig.DEFAULT_HOSTNAME
+ private var port: Int = 0
+
+ // Handle to the server
+ private var serverOpt: Option[NettyServer] = None
+
+ // The handler that handles the callback from Avro
+ private var handler: Option[SparkAvroCallbackHandler] = None
+
+ // Latch that blocks off the Flume framework from wasting 1 thread.
+ private val blockingLatch = new CountDownLatch(1)
+
+ override def start() {
+ handler = Option(new SparkAvroCallbackHandler(poolSize, getChannel,
transactionTimeout))
+ val responder = new SpecificResponder(classOf[SparkFlumeProtocol],
handler.get)
+ // Using the constructor that takes specific thread-pools requires
bringing in netty
+ // dependencies which are being excluded in the build. In practice,
+ // Netty dependencies are already available on the JVM as Flume would
have pulled them in.
+ serverOpt = Option(new NettyServer(responder, new
InetSocketAddress(hostname, port)))
+ serverOpt.map(server => {
+ server.start()
+ })
+ super.start()
+ }
+
+ override def stop() {
+ handler.map(callbackHandler => {
+ callbackHandler.shutdown()
+ })
+ serverOpt.map(server => {
+ server.close()
+ server.join()
+ })
+ blockingLatch.countDown()
+ super.stop()
+ }
+
+ /**
+ * @param ctx
+ */
+ override def configure(ctx: Context) {
+ import SparkSinkConfig._
+ hostname = ctx.getString(CONF_HOSTNAME, DEFAULT_HOSTNAME)
+ port = Option(ctx.getInteger(CONF_PORT)).
+ getOrElse(throw new ConfigurationException("The port to bind to must
be specified"))
+ poolSize = ctx.getInteger(THREADS, DEFAULT_THREADS)
+ transactionTimeout = ctx.getInteger(CONF_TRANSACTION_TIMEOUT,
DEFAULT_TRANSACTION_TIMEOUT)
+ }
+
+ override def process(): Status = {
+ // This method is called in a loop by the Flume framework - block it
until the sink is
+ // stopped to save CPU resources. The sink runner will interrupt this
thread when the sink is
+ // being shut down.
+ blockingLatch.await()
+ Status.BACKOFF
+ }
+}
+
+/**
+ * Class that implements the SparkFlumeProtocol, that is used by the Avro
Netty Server to process
+ * requests. Each getEvents, ack and nack call is forwarded to an instance
of this class.
+ * @param threads Number of threads to use to process requests.
+ * @param channel The channel that the sink pulls events from
+ * @param transactionTimeout Timeout in millis after which the transaction
if not acked by Spark
+ * is rolled back.
+ */
+private class SparkAvroCallbackHandler(val threads: Int, val channel:
Channel,
+ val transactionTimeout: Int) extends SparkFlumeProtocol {
+ private val LOG =
LoggerFactory.getLogger(classOf[SparkAvroCallbackHandler])
+ val transactionExecutorOpt = Option(Executors.newFixedThreadPool(threads,
+ new ThreadFactoryBuilder().setDaemon(true)
+ .setNameFormat("Spark Sink Processor Thread - %d").build()))
+ private val processorMap = new ConcurrentHashMap[CharSequence,
TransactionProcessor]()
+ // This sink will not persist sequence numbers and reuses them if it
gets restarted.
+ // So it is possible to commit a transaction which may have been meant
for the sink before the
+ // restart.
+ // Since the new txn may not have the same sequence number we must guard
against accidentally
+ // committing a new transaction. To reduce the probability of that
happening a random string is
+ // prepended to the sequence number. Does not change for life of sink
+ private val seqBase = RandomStringUtils.randomAlphanumeric(8)
+ private val seqCounter = new AtomicLong(0)
+
+ /**
+ * Returns a bunch of events to Spark over Avro RPC.
+ * @param n Maximum number of events to return in a batch
+ * @return [[EventBatch]] instance that has a sequence number and an
array of at most n events
+ */
+ override def getEventBatch(n: Int): EventBatch = {
+ val sequenceNumber = seqBase + seqCounter.incrementAndGet()
+ val processor = new TransactionProcessor(channel, sequenceNumber,
+ n, transactionTimeout, this)
+ transactionExecutorOpt.map(executor => {
+ executor.submit(processor)
+ })
+ // Wait until a batch is available - can be null if some error was
thrown
+ processor.getEventBatch match {
+ case ErrorEventBatch => throw new FlumeException("Something went
wrong. No events" +
+ " retrieved from channel.")
+ case eventBatch: EventBatch =>
+ processorMap.put(sequenceNumber, processor)
+ if (LOG.isDebugEnabled()) {
+ LOG.debug("Sent " + eventBatch.getEvents.size() +
+ " events with sequence number: " +
eventBatch.getSequenceNumber)
+ }
+ eventBatch
+ }
+ }
+
+ /**
+ * Called by Spark to indicate successful commit of a batch
+ * @param sequenceNumber The sequence number of the event batch that was
successful
+ */
+ override def ack(sequenceNumber: CharSequence): Void = {
+ completeTransaction(sequenceNumber, success = true)
+ null
+ }
+
+ /**
+ * Called by Spark to indicate failed commit of a batch
+ * @param sequenceNumber The sequence number of the event batch that
failed
+ * @return
+ */
+ override def nack(sequenceNumber: CharSequence): Void = {
+ completeTransaction(sequenceNumber, success = false)
+ LOG.info("Spark failed to commit transaction. Will reattempt events.")
+ null
+ }
+
+ /**
+ * Helper method to commit or rollback a transaction.
+ * @param sequenceNumber The sequence number of the batch that was
completed
+ * @param success Whether the batch was successful or not.
+ */
+ private def completeTransaction(sequenceNumber: CharSequence, success:
Boolean) {
+ Option(removeAndGetProcessor(sequenceNumber)).map(processor => {
+ processor.batchProcessed(success)
+ })
+ }
+
+ /**
+ * Helper method to remove the TxnProcessor for a Sequence Number. Can
be used to avoid a leak.
+ * @param sequenceNumber
+ * @return The transaction processor for the corresponding batch. Note
that this instance is no
+ * longer tracked and the caller is responsible for that txn
processor.
+ */
+ private[flume] def removeAndGetProcessor(sequenceNumber: CharSequence):
TransactionProcessor = {
+ processorMap.remove(sequenceNumber.toString) // The toString is
required!
+ }
+
+ /**
+ * Shuts down the executor used to process transactions.
+ */
+ def shutdown() {
+ transactionExecutorOpt.map(executor => {
+ executor.shutdownNow()
+ })
+ }
+}
+
+/**
+ * Object representing an empty batch returned by the txn processor due to
some error.
+ */
+case object ErrorEventBatch extends EventBatch
+
+// Flume forces transactions to be thread-local (horrible, I know!)
+// So the sink basically spawns a new thread to pull the events out within
a transaction.
+// The thread fills in the event batch object that is set before the
thread is scheduled.
+// After filling it in, the thread waits on a condition - which is
released only
+// when the success message comes back for the specific sequence number
for that event batch.
+/**
+ * This class represents a transaction on the Flume channel. This class
runs a separate thread
+ * which owns the transaction. The thread is blocked until the success
call for that transaction
+ * comes back with an ACK or NACK.
+ * @param channel The channel from which to pull events
+ * @param seqNum The sequence number to use for the transaction. Must be
unique
+ * @param maxBatchSize The maximum number of events to process per batch
+ * @param transactionTimeout Time in seconds after which a transaction
must be rolled back
+ * without waiting for an ACK from Spark
+ * @param parent The parent [[SparkAvroCallbackHandler]] instance, for
reporting timeouts
+ */
+private class TransactionProcessor(val channel: Channel, val seqNum:
String,
+ var maxBatchSize: Int, val transactionTimeout: Int,
+ val parent: SparkAvroCallbackHandler) extends Callable[Void] {
+
+ private val LOG = LoggerFactory.getLogger(classOf[TransactionProcessor])
+
+ // If a real batch is not returned, we always have to return an error
batch.
+ @volatile private var eventBatch: EventBatch = ErrorEventBatch
+
+ // Synchronization primitives
+ val batchGeneratedLatch = new CountDownLatch(1)
+ val batchAckLatch = new CountDownLatch(1)
+
+ // Sanity check to ensure we don't loop like crazy
+ val totalAttemptsToRemoveFromChannel = Int.MaxValue / 2
+
+ // OK to use volatile, since the change would only make this true
(otherwise it will be
+ // changed to false - we never apply a negation operation to this) -
which means the transaction
+ // succeeded.
+ @volatile private var batchSuccess = false
+
+ // The transaction that this processor would handle
+ var txOpt: Option[Transaction] = None
+
+ /**
+ * Get an event batch from the channel. This method will block until a
batch of events is
+ * available from the channel. If no events are available after a large
number of attempts of
+ * polling the channel, this method will return [[ErrorEventBatch]].
+ *
+ * @return An [[EventBatch]] instance with sequence number set to
[[seqNum]], filled with a
+ * maximum of [[maxBatchSize]] events
+ */
+ def getEventBatch: EventBatch = {
+ batchGeneratedLatch.await()
+ eventBatch
+ }
+
+ /**
+ * This method is to be called by the sink when it receives an ACK or
NACK from Spark. This
+ * method is a no-op if it is called after [[transactionTimeout]] has
expired since
+ * [[getEventBatch]] returned a batch of events.
+ * @param success True if an ACK was received and the transaction should
be committed, else false.
+ */
+ def batchProcessed(success: Boolean) {
+ if (LOG.isDebugEnabled) {
+ LOG.debug("Batch processed for sequence number: " + seqNum)
+ }
+ batchSuccess = success
+ batchAckLatch.countDown()
+ }
+
+ /**
+ * Populates events into the event batch. If the batch cannot be
populated,
+ * this method will not set the event batch which will stay
[[ErrorEventBatch]]
+ */
+ private def populateEvents() {
+ try {
+ txOpt = Option(channel.getTransaction)
+ txOpt.map(tx => {
+ tx.begin()
+ val events = new util.ArrayList[SparkSinkEvent](maxBatchSize)
+ val loop = new Breaks
+ var gotEventsInThisTxn = false
+ var loopCounter: Int = 0
+ loop.breakable {
+ while (events.size() < maxBatchSize
+ && loopCounter < totalAttemptsToRemoveFromChannel) {
+ loopCounter += 1
+ Option(channel.take()) match {
+ case Some(event) =>
+ events.add(new
SparkSinkEvent(toCharSequenceMap(event.getHeaders),
+ ByteBuffer.wrap(event.getBody)))
+ gotEventsInThisTxn = true
+ case None =>
+ if (!gotEventsInThisTxn) {
+ TimeUnit.MILLISECONDS.sleep(500)
+ } else {
+ loop.break()
+ }
+ }
+ }
+ }
+ if (!gotEventsInThisTxn) {
+ throw new FlumeException("Tried too many times, didn't get any
events from the channel")
+ }
+ // At this point, the events are available, so fill them into the
event batch
+ eventBatch = new EventBatch(seqNum, events)
+ })
+ } catch {
+ case e: Throwable =>
+ LOG.error("Error while processing transaction.", e)
+ try {
+ txOpt.map(tx => {
+ rollbackAndClose(tx, close = true)
+ })
+ } finally {
+ // Avro might serialize the exception and cause a NACK,
+ // so don't bother with the transaction
+ txOpt = None
+ }
+ } finally {
+ batchGeneratedLatch.countDown()
+ }
+ }
+
+ /**
+ * Waits for upto [[transactionTimeout]] seconds for an ACK. If an ACK
comes in,
+ * this method commits the transaction with the channel. If the ACK does
not come in within
+ * that time or a NACK comes in, this method rolls back the transaction.
+ */
+ private def processAckOrNack() {
+ batchAckLatch.await(transactionTimeout, TimeUnit.SECONDS)
+ txOpt.map(tx => {
+ if (batchSuccess) {
+ try {
+ tx.commit()
+ } catch {
+ case e: Throwable =>
+ rollbackAndClose(tx, close = false) // tx will be closed later
anyway
--- End diff --
The rollback logs it as a generic "rolling back" message; it does not log
the reason behind the rollback in detail enough. From what I see, there are a
number of cases when a rollback can occur
- no data received
- exception in receiving data
- data received, sent to receiver, but nack received
- data received, sent to receiver, but ack not received by timeout
- data receiver, sent to receiver, exception will waiting for ack (may be
same as above)
We should be able to very easily identify which of the cases caused the
rollback of the transaction. It need not require logging in the
rollbackAndClose function, but at least all control paths leading to a failure
of the transaction should get logged for easy debugging.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
