zhijiangW commented on a change in pull request #11515: [FLINK-16744][task] implement channel state persistence for unaligned checkpoints URL: https://github.com/apache/flink/pull/11515#discussion_r403487299
########## File path: flink-runtime/src/main/java/org/apache/flink/runtime/checkpoint/channel/ChannelStateWriterImpl.java ########## @@ -0,0 +1,225 @@ +/* + * 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.flink.runtime.checkpoint.channel; + +import org.apache.flink.annotation.Internal; +import org.apache.flink.runtime.checkpoint.CheckpointOptions; +import org.apache.flink.runtime.io.network.buffer.Buffer; +import org.apache.flink.runtime.state.CheckpointStorageWorkerView; +import org.apache.flink.runtime.state.CheckpointStreamFactory; +import org.apache.flink.util.Preconditions; + +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import javax.annotation.concurrent.ThreadSafe; + +import java.util.Map; +import java.util.concurrent.ArrayBlockingQueue; +import java.util.concurrent.BlockingQueue; +import java.util.concurrent.ConcurrentHashMap; + +import static org.apache.flink.runtime.checkpoint.channel.ChannelStateWriteRequest.completeInput; +import static org.apache.flink.runtime.checkpoint.channel.ChannelStateWriteRequest.completeOutput; +import static org.apache.flink.runtime.checkpoint.channel.ChannelStateWriteRequest.write; + +/** + * {@link ChannelStateWriter} implemented using + * {@link CheckpointStreamFactory.CheckpointStateOutputStream CheckpointStateOutputStreams}. Internally, it has + * <ul> + * <li>one stream per checkpoint; having multiple streams would mean more files written and more connections opened + * (and more latency on restore)</li> + * <li>one thread; having multiple threads means more connections, couples with the implementation and increases complexity</li> + * </ul> + */ +@Internal +@ThreadSafe +public class ChannelStateWriterImpl implements ChannelStateWriter { + + private static final Logger LOG = LoggerFactory.getLogger(ChannelStateWriterImpl.class); + private static final int DEFAULT_HANDOVER_CAPACITY = 10; + private static final int DEFAULT_MAX_CHECKPOINTS = 5; // currently, only single in-flight checkpoint is supported + + private class ProcessRequestsLoop implements Runnable { + private final ChannelStateWriteRequestProcessor requestProcessor; + + private ProcessRequestsLoop(int maxCheckpoints, CheckpointStorageWorkerView streamFactoryFactory, ChannelStateSerializer serializer) { + this.requestProcessor = new ChannelStateWriteRequestProcessor(maxCheckpoints, streamFactoryFactory, serializer); + } + + @Override + public void run() { + try { + loop(); + } catch (Exception ex) { + thrown = ex; + } finally { + handover.clear(); + requestProcessor.cleanup(thrown != null ? thrown : new RuntimeException("loop terminated, isRunning: " + isRunning)); + } + LOG.debug("loop terminated"); + } + + private void loop() throws Exception { + while (isRunning || !handover.isEmpty()) { + try { + requestProcessor.processRequest(handover.take()); + } catch (InterruptedException e) { + if (isRunning || !handover.isEmpty()) { + LOG.info("interrupted while waiting for an item (continue waiting)", e); + } else { + Thread.currentThread().interrupt(); + return; + } + } + } + } + } + + private final Thread asyncWriter; + private final BlockingQueue<ChannelStateWriteRequest> handover; Review comment: Thanks for sharing your thoughts. Actually I only thought of the benefits of resource limitation for bounded queue before. Do you concerning that the network buffer to be exhausted or run out of memory by unbounded queue I f I understand correctly? If so, I guess this hurt might be less worse than blocking the netty thread for two reasons: - If one netty thread is blocked, it might block multiple input channels receiving following data, then it might even delay the barrier alignment, based on the truth that one netty thread wouold serve for multiple input channels data transport. - If one netty thread is blocked, it does not alway mean that the left buffer memory can be better used by other channels. For exclusive buffers per-channel, it can only be used by current blocked netty thread for receiving data. For floating buffers among all the channels, if they were already requested away by the current blocked netty thread, they can not returned back to be used by other active netty threads. It can only avoid requesting more floating buffers by current blocked netty thread, but it is not determined behavior. So if there are no concerns for additional memory overhead caused by unbounded queue and it is also easy to adjust, then i prefer to unblocking option. Otherwise we can lift the default size for bounded size and further convert to unbounded queue if finding serious problems in production future. But I am not quite sure whether the blocking impact can be easily found in practice. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
