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:
- single:每个算子单独统计延迟;
- operator(默认值):每个下游算子都统计自己与Source算子之间的延迟;
- subtask:每个下游算子的sub-task都统计自己与Source算子的sub-task之间的延迟。
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:
- ProcessingTimeService产生时间戳最终是靠
System.currentTimeMillis()
方法,所以必须保证Flink集群内所有节点的时区、时间是同步的,可以用ntp等工具来配置。 -
metrics.latency.interval
的时间间隔宜大不宜小,在我们的实践中一般配置成30000(30秒)左右。一是因为延迟监控的频率可以不用太频繁,二是因为LatencyMarker的处理也要消耗时间,只有在LatencyMarker的耗时远小于正常StreamRecord的耗时时,metric反映出的数据才贴近实际情况,所以LatencyMarker的密度不能太大。
The End
待会该买菜做饭了,就这样吧。
民那周末愉快(不是