Flink

Flink链路延迟测量的LatencyMarker机制实现

2020-06-07  本文已影响0人  LittleMagic

前言

今天本应放一首适合高考气氛的歌的，但是既然受疫情影响推迟了，还是老老实实写点技术相关的吧。

对于实时的流式处理系统来说，我们需要关注数据输入、计算和输出的及时性，所以处理延迟是一个比较重要的监控指标，特别是在数据量大或者软硬件条件不佳的环境下。Flink早在FLINK-3660就为用户提供了开箱即用的链路延迟测量功能，只需要配置好metrics.latency.interval参数，再观察TaskManagerJobMetricGroup/operator_id/operator_subtask_index/latency这个metric即可。本文简单walk一下源码，看看它是如何实现的，并且简要说明注意事项。

LatencyMarker的产生

与通过水印来标记事件时间的推进进度相似，Flink也用一种特殊的流元素（StreamElement）作为延迟的标记，称为LatencyMarker。

LatencyMarker的数据结构甚简单，只有3个field，即它被创建时携带的时间戳、算子ID和算子并发实例（sub-task）的ID。

private final long markedTime;
private final OperatorID operatorId;
private final int subtaskIndex;

LatencyMarker和水印不同，不需要通过用户抽取产生，而是在Source端自动按照metrics.latency.interval参数指定的周期生成。StreamSource专门实现了一个内部类LatencyMarksEmitter用来发射LatencyMarker，而它又借用了负责协调处理时间的服务ProcessingTimeService（之前的文章已经多次提到过），如下代码所示。

LatencyMarksEmitter<OUT> latencyEmitter = null;
if (latencyTrackingInterval > 0) {
    latencyEmitter = new LatencyMarksEmitter<>(
        getProcessingTimeService(),
        collector,
        latencyTrackingInterval,
        this.getOperatorID(),
        getRuntimeContext().getIndexOfThisSubtask());
}

private static class LatencyMarksEmitter<OUT> {
    private final ScheduledFuture<?> latencyMarkTimer;

    public LatencyMarksEmitter(
            final ProcessingTimeService processingTimeService,
            final Output<StreamRecord<OUT>> output,
            long latencyTrackingInterval,
            final OperatorID operatorId,
            final int subtaskIndex) {
        latencyMarkTimer = processingTimeService.scheduleAtFixedRate(
            new ProcessingTimeCallback() {
                @Override
                public void onProcessingTime(long timestamp) throws Exception {
                    try {
                        // ProcessingTimeService callbacks are executed under the checkpointing lock
                        output.emitLatencyMarker(new LatencyMarker(processingTimeService.getCurrentProcessingTime(), operatorId, subtaskIndex));
                    } catch (Throwable t) {
                        // we catch the Throwables here so that we don't trigger the processing
                        // timer services async exception handler
                        LOG.warn("Error while emitting latency marker.", t);
                    }
                }
            },
            0L,
            latencyTrackingInterval);
    }

    public void close() {
        latencyMarkTimer.cancel(true);
    }
}

通过调用Output.emitLatencyMarker()方法，LatencyMarker就会随着数据流一起传递到下游了。

LatencyMarker的粒度

AbstractStreamOperator是所有Flink Streaming算子的基类，在它的初始化方法setup()中，会先创建用于延迟统计的LatencyStats实例。

final String configuredGranularity = taskManagerConfig.getString(MetricOptions.LATENCY_SOURCE_GRANULARITY);
LatencyStats.Granularity granularity;
try {
    granularity = LatencyStats.Granularity.valueOf(configuredGranularity.toUpperCase(Locale.ROOT));
} catch (IllegalArgumentException iae) {
    granularity = LatencyStats.Granularity.OPERATOR;
    LOG.warn(
        "Configured value {} option for {} is invalid. Defaulting to {}.",
        configuredGranularity,
        MetricOptions.LATENCY_SOURCE_GRANULARITY.key(),
        granularity);
}
TaskManagerJobMetricGroup jobMetricGroup = this.metrics.parent().parent();
this.latencyStats = new LatencyStats(jobMetricGroup.addGroup("latency"),
    historySize,
    container.getIndexInSubtaskGroup(),
    getOperatorID(),
    granularity);

在创建LatencyStats之前，先要根据metrics.latency.granularity配置项来确定延迟监控的粒度，分为以下3档：

一般情况下采用默认的operator粒度即可，这样在Sink端观察到的latency metric就是我们最想要的全链路（端到端）延迟，以下也是以该粒度讲解。subtask粒度太细，会增大所有并行度的负担，不建议使用。

LatencyMarker的流转与计量

AbstractStreamOperator分别提供了用于单输入流算子OneInputStreamOperator、双输入流算子TwoInputStreamOperator的LatencyMarker处理方法。

// ------- One input stream
public void processLatencyMarker(LatencyMarker latencyMarker) throws Exception {
    reportOrForwardLatencyMarker(latencyMarker);
}

// ------- Two input stream
public void processLatencyMarker1(LatencyMarker latencyMarker) throws Exception {
    reportOrForwardLatencyMarker(latencyMarker);
}

public void processLatencyMarker2(LatencyMarker latencyMarker) throws Exception {
    reportOrForwardLatencyMarker(latencyMarker);
}

protected void reportOrForwardLatencyMarker(LatencyMarker marker) {
    // all operators are tracking latencies
    this.latencyStats.reportLatency(marker);
    // everything except sinks forwards latency markers
    this.output.emitLatencyMarker(marker);
}

这些方法都会做两件事，一是计算延时并报告给LatencyStats，二是继续将LatencyMarker发射到下游。不妨来看看RecordWriterOutput.emitLatencyMarker()方法的具体实现。

@Override
public void emitLatencyMarker(LatencyMarker latencyMarker) {
    serializationDelegate.setInstance(latencyMarker);
    try {
        recordWriter.randomEmit(serializationDelegate);
    }
    catch (Exception e) {
        throw new RuntimeException(e.getMessage(), e);
    }
}

/**
 * This is used to send LatencyMarks to a random target channel.
 */
public void randomEmit(T record) throws IOException, InterruptedException {
    emit(record, rng.nextInt(numberOfChannels));
}

可见是从该算子所有的输出channel中随机选择一条来发射LatencyMarker，这样在度量算子级别延迟的基础上不会造成LatencyMarker泛滥，同时也不会受到并行度调整（重新分区）的影响。

注意StreamSink的reportOrForwardLatencyMarker()方法不会再发射LatencyMarker（因为已经处理完了），只会更新延迟。

@Override
protected void reportOrForwardLatencyMarker(LatencyMarker marker) {
    // all operators are tracking latencies
    this.latencyStats.reportLatency(marker);
    // sinks don't forward latency markers
}

LatencyStats中的延迟最终会转化为直方图表示，通过直方图就可以统计出延时的最大值、最小值、均值、分位值（quantile）等指标。以下是reportLatency()方法的源码。

public void reportLatency(LatencyMarker marker) {
    final String uniqueName = granularity.createUniqueHistogramName(marker, operatorId, subtaskIndex);
    DescriptiveStatisticsHistogram latencyHistogram = this.latencyStats.get(uniqueName);
    if (latencyHistogram == null) {
        latencyHistogram = new DescriptiveStatisticsHistogram(this.historySize);
        this.latencyStats.put(uniqueName, latencyHistogram);
        granularity.createSourceMetricGroups(metricGroup, marker, operatorId, subtaskIndex)
            .addGroup("operator_id", String.valueOf(operatorId))
            .addGroup("operator_subtask_index", String.valueOf(subtaskIndex))
            .histogram("latency", latencyHistogram);
    }
    long now = System.currentTimeMillis();
    latencyHistogram.update(now - marker.getMarkedTime());
}

可见，延迟是由当前时间戳减去LatencyMarker携带的时间戳得到的，所以在Sink端统计到的就是全链路延迟了。

注意事项

由以上分析可知，LatencyMarker是不会像Watermark一样参与到数据流的用户逻辑中的，而是直接被各算子转发并统计。这如何能得到真正的延时呢？如果由于网络不畅、数据流量太大等原因造成了反压（back pressure，之后再提），那么LatencyMarker的流转就会被阻碍，传递到下游的时间差就会增加，所以还是能够近似估算出整体的延时的。为了让它尽量精确，有两点特别需要注意：

The End

待会该买菜做饭了，就这样吧。

民那周末愉快（不是

上一篇下一篇

猜你喜欢

热点阅读