flink key 通过添加延迟；我怎样才能减少这种延迟？

Question

当我使用 KeyedStream 运行一个简单的 flink 应用程序时，我观察到一个事件的时间延迟从 0 到 100 毫秒不等。下面是程序

StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
    DataStream<Long> source = env.addSource(new SourceFunction<Long>() {
        public void run(SourceContext<Long> sourceContext) throws Exception {
            while(true) {
                synchronized (sourceContext.getCheckpointLock()) {
                    sourceContext.collect(System.currentTimeMillis());
                    Thread.sleep(1000);
                }
            }
        }

        public void cancel() {}
    }).keyBy(new KeySelector<Long, Long>() {
        @Override
        public Long getKey(Long l) throws Exception {
            return l;
        }
    }).addSink(new SinkFunction<Long>() {
        @Override
        public void invoke(Long l) throws Exception {
            long diff = System.currentTimeMillis() - l;
            System.out.println("in Sink: diff=" + diff);
        }
    });
    env.execute();

输出是：

in Sink: diff=0
in Sink: diff=2
in Sink: diff=4
in Sink: diff=4
in Sink: diff=5
in Sink: diff=7
in Sink: diff=9
in Sink: diff=9
in Sink: diff=11
in Sink: diff=12
in Sink: diff=14
in Sink: diff=14
in Sink: diff=16
in Sink: diff=17
in Sink: diff=18
in Sink: diff=19
in Sink: diff=21
in Sink: diff=22
in Sink: diff=24
in Sink: diff=24
in Sink: diff=26
in Sink: diff=27
in Sink: diff=29
in Sink: diff=29
in Sink: diff=31
in Sink: diff=32
in Sink: diff=34
in Sink: diff=34
in Sink: diff=36
in Sink: diff=37
in Sink: diff=39
in Sink: diff=40
in Sink: diff=41
in Sink: diff=43
in Sink: diff=45
in Sink: diff=45
in Sink: diff=47
in Sink: diff=48
in Sink: diff=50
in Sink: diff=50
in Sink: diff=52
in Sink: diff=53
in Sink: diff=55
in Sink: diff=57
in Sink: diff=57
in Sink: diff=59
in Sink: diff=60
in Sink: diff=61
in Sink: diff=62
in Sink: diff=63
in Sink: diff=65
in Sink: diff=66
in Sink: diff=67
in Sink: diff=69
in Sink: diff=70
in Sink: diff=72
in Sink: diff=72
in Sink: diff=74
in Sink: diff=76
in Sink: diff=77
in Sink: diff=78
in Sink: diff=79
in Sink: diff=81
in Sink: diff=82
in Sink: diff=83
in Sink: diff=84
in Sink: diff=86
in Sink: diff=87
in Sink: diff=88
in Sink: diff=89
in Sink: diff=91
in Sink: diff=92
in Sink: diff=94
in Sink: diff=94
in Sink: diff=96
in Sink: diff=97
in Sink: diff=99
in Sink: diff=99
in Sink: diff=0
in Sink: diff=2
in Sink: diff=3
in Sink: diff=4
in Sink: diff=4
in Sink: diff=5
in Sink: diff=7
in Sink: diff=9
in Sink: diff=9
in Sink: diff=11
in Sink: diff=12
in Sink: diff=14
in Sink: diff=14
in Sink: diff=16
in Sink: diff=17
in Sink: diff=18
in Sink: diff=19
in Sink: diff=21
in Sink: diff=22
in Sink: diff=24
in Sink: diff=24
in Sink: diff=26
in Sink: diff=46
in Sink: diff=48
in Sink: diff=50
in Sink: diff=52
in Sink: diff=53
in Sink: diff=54
in Sink: diff=56
in Sink: diff=58
in Sink: diff=59
in Sink: diff=60
in Sink: diff=62
in Sink: diff=64
in Sink: diff=65
in Sink: diff=66
in Sink: diff=68
in Sink: diff=70
in Sink: diff=71
in Sink: diff=73
in Sink: diff=74
in Sink: diff=76
in Sink: diff=77
in Sink: diff=79
in Sink: diff=81
in Sink: diff=82
in Sink: diff=83
in Sink: diff=85
in Sink: diff=86
in Sink: diff=88
in Sink: diff=88
in Sink: diff=90
in Sink: diff=92
in Sink: diff=92
in Sink: diff=94
in Sink: diff=95
in Sink: diff=97
in Sink: diff=98
in Sink: diff=99
in Sink: diff=0
in Sink: diff=2
in Sink: diff=4
in Sink: diff=4
in Sink: diff=5
in Sink: diff=7
in Sink: diff=9

如您所见，延迟是一种模式，逐渐增加到 100，下降并从 0 开始，循环重复。我需要尽可能低的延迟。这个例子是我真实应用程序的简化版本。有人可以解释一下延迟的原因以及如何将延迟降低到尽可能低。

Answer 1

这种延迟的原因是，通过添加该 keyBy，您将强制进行网络洗牌以及序列化/反序列化。延迟如此多变的原因是网络缓冲。

您需要阅读文档中名为Controlling Latency 的部分。 tl;dr 是您想将网络缓冲区超时设置为较小的值：

env.setBufferTimeout(timeoutMillis);

如果需要，您可以将缓冲区超时设置为零，但这对吞吐量的影响大于将其设置为较小的值（如 1 毫秒或 5 毫秒）。默认值为 100 毫秒。 Flink 中网络栈的组织方式详见 Flink 项目博客A Deep-Dive into Flink's Network Stack。

在我们讨论这个主题时，其他延迟来源可能包括检查点屏障对齐和垃圾收集。

env.getCheckpointConfig().setCheckpointingMode(CheckpointingMode.AT_LEAST_ONCE);

将禁用屏障对齐，代价是放弃一次处理语义。

使用 RocksDB 状态后端将减少垃圾收集的对象数量（因为 RocksDB 将其状态保持在堆外），在某些情况下会以更糟糕的平均延迟为代价来改善最坏情况的延迟。然而，现代垃圾收集器使用 RocksDB 来改善最坏情况下的延迟可能是错误的。

还有，

env.getConfig().enableObjectReuse();

将指示运行时重用用户对象以获得更好的性能。请记住，当用户代码函数不知道这种行为时，这可能会导致错误。

如果您使用水印，水印延迟会影响触发事件时间计时器的延迟（包括窗口），autoWatermarkInterval 也会影响延迟。

最后，事务接收器的使用增加了端到端延迟，因为这些接收器的下游消费者在事务完成之前不会看到提交的结果。预期的延迟大约是检查点间隔的一半。

如果您对测量延迟感兴趣，请查看Latency Tracking 和Monitoring Apache Flink Applications 101 中关于延迟的部分。