今天我们来讲讲Tracing系统中的时钟。目前市面上比较火的几款开源的链路追踪产品主要有
那么在链路追踪中，需要在客户端记录Span的起始时间点和持续时间，在Zipkin中，在Span中可以找到如下代码，
package zipkin2;
public final class Span implements Serializable { // for Spark and Flink jobs
<SNIP>
final long timestamp, duration; // zero means null, saving 2 object references
<SNIP>
}timestamp和duration分别记录Span的起始时间点和持续时间。类似得，在JaegerSpan中有
package io.jaegertracing.internal;
public class JaegerSpan implements Span {
...
private final long startTimeMicroseconds;
private final long startTimeNanoTicks;
private long durationMicroseconds; // span durationMicroseconds
...
}startTime*以及durationMicroseconds分别表示Span的起始时间点和持续时间，并且这里清楚地表明了记录的时间单位是微秒(microsecond)。在Skywalking中则为，AbstractTracingSpan。
package org.apache.skywalking.apm.agent.core.context.trace;
public abstract class AbstractTracingSpan implements AbstractSpan {
/**
* The start time of this Span.
*/
protected long startTime;
/**
* The end time of this Span.
*/
protected long endTime;
}其中的startTime以及endTime对应起始时间和结束时间，这里Skywalking用的并不是duration。
在Zipkin的客户端实现中，Brave使用了一种比较特别的方式。还记得我们上次在讲PendingSpans得时候留下的坑吗?
package brave.internal.recorder;
public final class PendingSpans extends ReferenceQueue<TraceContext> {
public PendingSpan getOrCreate(TraceContext context, boolean start) {
if (context == null) throw new NullPointerException("context == null");
reportOrphanedSpans();
PendingSpan result = delegate.get(context);
if (result != null) return result;
MutableSpan data = new MutableSpan();
if (context.shared()) data.setShared();
// 通常在创建一个新的Spand的时候，他的parentSpan应该会在执行状态
// 那么Brave为了节约计算时间的额外损耗，这里会首先获取这个context的父级
TickClock clock = getClockFromParent(context);
// 如果无法获取到父级，一般可能是这是一个新的Span，或者是父级的Span已经被回收了
if (clock == null) {
clock = new TickClock(this.clock.currentTimeMicroseconds(), System.nanoTime());
if (start) data.startTimestamp(clock.baseEpochMicros);
} else if (start) {
data.startTimestamp(clock.currentTimeMicroseconds());
}
PendingSpan newSpan = new PendingSpan(data, clock);
PendingSpan previousSpan = delegate.putIfAbsent(new RealKey(context, this), newSpan);
if (previousSpan != null) return previousSpan; // lost race
if (trackOrphans) {
newSpan.caller =
new Throwable("Thread " + Thread.currentThread().getName() + " allocated span here");
}
return newSpan;
}
}这里的TickClock是很有讲究的，一般在JDK9以前，Java中是无法获取到微秒级别的绝对时间精度的。比如在JDK8中，通过System.nanoTime()得到的时间只能用于计算相对时间，它的返回值并不能与任何真实时间挂钩。比如Oracle官方给出的案例，
long startTime = System.nanoTime();
// ... the code being measured ...
long estimatedTime = System.nanoTime() - startTime;可以被用于计算相对时间，能够达到纳秒的精度。而这里Brave采用了一个非常巧妙的方式，从TickClock中可以看到，
package brave.internal.recorder;
import brave.Clock;
final class TickClock implements Clock {
final long baseEpochMicros;
final long baseTickNanos;
TickClock(long baseEpochMicros, long baseTickNanos) {
// 基准绝对时间，单位us
this.baseEpochMicros = baseEpochMicros;
// 基准相对时间，单位ns
this.baseTickNanos = baseTickNanos;
}
@Override public long currentTimeMicroseconds() {
// 在计算当前时间的时候，会通过当前的nanoTime() - 基准相对时间，
// 再加上 基准绝对时间 就可以计算得到当前的时间，其中流逝的时间精度为纳秒
return ((System.nanoTime() - baseTickNanos) / 1000) + baseEpochMicros;
}
@Override public String toString() {
return "TickClock{"
+ "baseEpochMicros=" + baseEpochMicros + ", "
+ "baseTickNanos=" + baseTickNanos
+ "}";
}
}由于这里流逝的时间精度为纳秒，可以很好的满足Span中Duration计时的时间精度。但这里baseEpochMicros的时间精度则是取决于JDK的版本。
我们首先看一下Brave中Clock这个接口，
package brave;
// FunctionalInterface except Java language level 6
public interface Clock {
// 是一个FunctionalInterface，只有一个方法
long currentTimeMicroseconds();
}Brave中把一些平台相关的功能都放到了Platform这个类里面。
package brave.internal;
public abstract class Platform {
...
public Clock clock() {
return new Clock() {
// <= JDK8
@Override public long currentTimeMicroseconds() {
return System.currentTimeMillis() * 1000;
}
@Override public String toString() {
return "System.currentTimeMillis()";
}
};
}
static class Jre9 extends Jre7 {
@IgnoreJRERequirement @Override public Clock clock() {
// JDK9+
return new Clock() {
// we could use jdk.internal.misc.VM to do this more efficiently, but it is internal
@Override public long currentTimeMicroseconds() {
java.time.Instant instant = java.time.Clock.systemUTC().instant();
return (instant.getEpochSecond() * 1000000) + (instant.getNano() / 1000);
}
@Override public String toString() {
return "Clock.systemUTC().instant()";
}
};
}
@Override public String toString() {
return "Jre9{}";
}
}
}System.currentTimeMillis() * 1000，则精度为毫秒Instant来计算得到微秒精度的时间。The range of an instant requires the storage of a number larger than a long. To achieve this, the class stores a long representing epoch-seconds and an int representing nanosecond-of-second, which will always be between 0 and 999,999,999. The epoch-seconds are measured from the standard Java epoch of 1970-01-01T00:00:00Z where instants after the epoch have positive values, and earlier instants have negative values. For both the epoch-second and nanosecond parts, a larger value is always later on the time-line than a smaller value.
根据官方文档，Instant使用的是epoch-seconds+nanosecond-of-second来表示当前时间，可以达到纳秒的精度。如此也就可以理解上述的TickClock了。
我们在刚才看到Jaeger的代码，发现他有两个startTime的表示
要理解上述的定义需要先看一下Jaeger中对Clock接口的定义，
package io.jaegertracing.internal.clock;
/**
* A small abstraction around system clock that aims to provide microsecond precision with the best
* accuracy possible.
*/
public interface Clock {
/**
* Returns the current time in microseconds.
*
* @return the difference, measured in microseconds, between the current time and and the Epoch
* (that is, midnight, January 1, 1970 UTC).
*/
long currentTimeMicros();
/**
* Returns the current value of the running Java Virtual Machine's high-resolution time source, in
* nanoseconds.
*
* <p>
* This method can only be used to measure elapsed time and is not related to any other notion of
* system or wall-clock time.
*
* @return the current value of the running Java Virtual Machine's high-resolution time source, in
* nanoseconds
*/
long currentNanoTicks();
/**
* @return true if the time returned by {@link #currentTimeMicros()} is accurate enough to
* calculate span duration as (end-start). If this method returns false, the {@code JaegerTracer} will
* use {@link #currentNanoTicks()} for calculating duration instead.
*/
boolean isMicrosAccurate();
}这里的文档写得很清楚，比较特别的一个接口方法是boolean isMicrosAccurate()
currentTimeMicros()返回的时间足够精确，则为TrueFalse。Jaeger会转而使用currentNanoTicks()来计算相对时间。默认的SystemClock就是一个不精确的例子在JaegerTracer中可以看到具体对时间赋值的代码，
package io.jaegertracing.internal;
public class JaegerTracer implements Tracer, Closeable {
...
public class SpanBuilder implements Tracer.SpanBuilder {
@Override
public JaegerSpan start() {
...
long startTimeNanoTicks = 0;
boolean computeDurationViaNanoTicks = false;
// 如果用户为特意指定开始时间
if (startTimeMicroseconds == 0) {
// 使用默认的currentTimeMicros
startTimeMicroseconds = clock.currentTimeMicros();
// 接着检查时钟是否足够精确
if (!clock.isMicrosAccurate()) {
// 如果不够精确，则记录当前的相对纳秒数
startTimeNanoTicks = clock.currentNanoTicks();
computeDurationViaNanoTicks = true;
}
}
...
}
}
}在Span中记录了startTimeNanoTicks并设置computeDurationViaNanoTicks为True，则在JaegerSpan.finish()的时候会使用NanoTicks来计算duration。
package io.jaegertracing.internal;
public class JaegerSpan implements Span {
...
@Override
public void finish() {
if (computeDurationViaNanoTicks) {
long nanoDuration = tracer.clock().currentNanoTicks() - startTimeNanoTicks;
finishWithDuration(nanoDuration / 1000);
} else {
finish(tracer.clock().currentTimeMicros());
}
}
...
}这个代码应该足够简单了，不需要过多的解释。
精度是毫秒，直接用的System.currentTimeMillis()，具体可以看代码AbstractTracingSpan。在与其他两者的比较中落了下乘。
我们发现，老牌的Tracing系统Zipkin/Brave对客户端时间记录的抽象程度、时间精度和性能考量都十分到位。而Jaeger的抽象则稍有些繁琐，到时间精度的实现也能够达到微秒量级。但Skywalking则为对此做特殊优化。
从这方面考虑，Zipkin >= Jaeger » Skywalking
Copyright© 2013-2020
All Rights Reserved 京ICP备2023019179号-8