flink-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From GitBox <...@apache.org>
Subject [GitHub] zhijiangW commented on a change in pull request #6417: [FLINK-9913][runtime] Improve output serialization only once in RecordWriter
Date Tue, 28 Aug 2018 15:06:21 GMT
zhijiangW commented on a change in pull request #6417: [FLINK-9913][runtime] Improve output
serialization only once in RecordWriter
URL: https://github.com/apache/flink/pull/6417#discussion_r213343581
 
 

 ##########
 File path: flink-runtime/src/main/java/org/apache/flink/runtime/io/network/api/writer/RecordWriter.java
 ##########
 @@ -87,62 +86,71 @@ public RecordWriter(ResultPartitionWriter writer, ChannelSelector<T>
channelSele
 
 		this.numChannels = writer.getNumberOfSubpartitions();
 
-		/*
-		 * The runtime exposes a channel abstraction for the produced results
-		 * (see {@link ChannelSelector}). Every channel has an independent
-		 * serializer.
-		 */
-		this.serializers = new SpanningRecordSerializer[numChannels];
+		this.serializer = new SpanningRecordSerializer<T>();
 		this.bufferBuilders = new Optional[numChannels];
 		for (int i = 0; i < numChannels; i++) {
-			serializers[i] = new SpanningRecordSerializer<T>();
 			bufferBuilders[i] = Optional.empty();
 		}
 	}
 
 	public void emit(T record) throws IOException, InterruptedException {
+		serializer.serializeRecord(record);
+
 		for (int targetChannel : channelSelector.selectChannels(record, numChannels)) {
-			sendToTarget(record, targetChannel);
+			copyToTarget(targetChannel);
 		}
+
+		// Make sure we don't hold onto the large intermediate serialization buffer for too long
+		serializer.prune();
 	}
 
 	/**
 	 * This is used to broadcast Streaming Watermarks in-band with records. This ignores
 	 * the {@link ChannelSelector}.
 	 */
 	public void broadcastEmit(T record) throws IOException, InterruptedException {
+		serializer.serializeRecord(record);
+
 		for (int targetChannel = 0; targetChannel < numChannels; targetChannel++) {
-			sendToTarget(record, targetChannel);
+			copyToTarget(targetChannel);
 		}
+
+		serializer.prune();
 	}
 
 	/**
 	 * This is used to send LatencyMarks to a random target channel.
 	 */
 	public void randomEmit(T record) throws IOException, InterruptedException {
-		sendToTarget(record, rng.nextInt(numChannels));
-	}
+		serializer.serializeRecord(record);
 
-	private void sendToTarget(T record, int targetChannel) throws IOException, InterruptedException
{
-		RecordSerializer<T> serializer = serializers[targetChannel];
+		copyToTarget(rng.nextInt(numChannels));
 
-		SerializationResult result = serializer.addRecord(record);
+		serializer.prune();
+	}
 
+	private void copyToTarget(int targetChannel) throws IOException, InterruptedException {
+		// We should reset the initial position of the intermediate serialization buffer before
+		// copying, so the serialization results can be copied to multiple target buffers.
+		serializer.reset();
+
+		BufferBuilder bufferBuilder = getBufferBuilder(targetChannel);
+		SerializationResult result = serializer.copyToBufferBuilder(bufferBuilder);
 		while (result.isFullBuffer()) {
-			if (tryFinishCurrentBufferBuilder(targetChannel, serializer)) {
-				// If this was a full record, we are done. Not breaking
-				// out of the loop at this point will lead to another
-				// buffer request before breaking out (that would not be
-				// a problem per se, but it can lead to stalls in the
-				// pipeline).
-				if (result.isFullRecord()) {
-					break;
-				}
+			tryFinishCurrentBufferBuilder(targetChannel);
 
 Review comment:
   You pointed out a good question!
   1. Considering `tryFinishCurrentBufferBuilder()`, the logic is somewhat different from
before. In the past, the buffer builder may be empty when calling `tryFinishCurrentBufferBuilder()`,
then it returns a boolean value to indicate the result. But now, we know the buffer builder
is always present when calling `tryFinishCurrentBufferBuilder`, so we may change it to `finishCurrentBufferBuilder()`
seems more appropriate. And adds the check code instead as following:
     ```
   private void finishCurrentBufferBuilder(int targetChannel) {
   		checkState(bufferBuilders[targetChannel].isPresent());
   
   		BufferBuilder bufferBuilder = bufferBuilders[targetChannel].get();
   		bufferBuilders[targetChannel] = Optional.empty();
   		numBytesOut.inc(bufferBuilder.finish());
   	}
   ```
   2. But I think we still need call `finishCurrentBufferBuilder()`  after checking every
full buffer. Because in the while loop the serialized record may be copied into several different
buffer builders, and each buffer build referenced by the `bufferBuilders ` needs to be finished
when full.
   
   3. It actually seems better than before, as the loop exit not rely on the return value
of `tryFinishCurrentBufferBuilder()`.

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services

Mime
View raw message