最近在研究全链路跟踪，因某些原因选用了 ZipKin 的 Brave 作为埋点工具，ZipKin不使用，本系列仅做 Brave(版本=5.6.0) 部分源码解读，其他内容不涉及。
对全链路不熟悉的，先学习 opentracing 的标准规范：https://opentracing-contrib.github.io/opentracing-specification-zh/specification.html
先来看一个简单埋点Demo：
import brave.Span;
import brave.Tracer;
import brave.Tracer.SpanInScope;
import brave.Tracing;
import brave.propagation.B3Propagation;
import brave.propagation.ExtraFieldPropagation;
public class Atest {
public static void main(String[] args) {
Tracing tracing = Tracing.newBuilder().localServiceName("test")
.propagationFactory(ExtraFieldPropagation.newFactory(B3Propagation.FACTORY, "user-name"))
.build();
Tracer tracer = tracing.tracer();
Span root = tracer.nextSpan().name("root").start();//创建跨度 root
SpanInScope scope = null;
try {
scope = tracer.withSpanInScope(root);//设置 root 跨度的作用域
//开始新的跨度 s1。此处使用 currentTraceContext().get() 获取到当前作用域中的 TraceContext
//TraceContext 中包含着链路中的关键信息，如 TraceId, parentId, spanId 等
Span s1 = tracer.newChild(tracing.currentTraceContext().get()).name("s1").start();
System.out.println("被跟踪的业务代码...");
s1.finish();//结束跨度 s1
} catch (Exception e) {
root.error(e);//报错处理
} finally {
scope.close();//结束作用域
}
root.finish();//结束跨度 root
}
}
执行代码，使用的是 Brave 默认的 LoggerReporter，输出的日志如下，默认使用的是V2版本，比起V1版本的日志，节省了很多日志内容，更加简洁易懂且提升性能。
可见“被跟踪的业务代码...”先被打印出来，接着打印的是跨度 s1，它的父节点是“b236c6f732bca43f”，spanId是“a41bf8dd8fba3f16”，耗时 138 微秒，serviceName是在Tracing设置的“test”，ip地址是“192.168.1.6”
最后打印的是跨度 root，一个完整简单的链路跟踪就完成了。
被跟踪的业务代码...
一月 24, 2019 10:20:11 下午 brave.Tracing$LoggingReporter report
信息: {"traceId":"b236c6f732bca43f","parentId":"b236c6f732bca43f","id":"a41bf8dd8fba3f16","name":"s1","timestamp":1548339611584476,"duration":138,"localEndpoint":{"serviceName":"test","ipv4":"192.168.1.6"}}
一月 24, 2019 10:20:11 下午 brave.Tracing$LoggingReporter report
信息: {"traceId":"b236c6f732bca43f","id":"b236c6f732bca43f","name":"root","timestamp":1548339611569087,"duration":53153,"localEndpoint":{"serviceName":"test","ipv4":"192.168.1.6"}}
本节先看 Tracing 的代码节选，Tracing 主要用于初始化链路跟踪所需的各组件，使用了 Builder 的模式，可根据需求自由去创建合适的链路跟踪特性。
譬如上面所使用的默认的 LoggingReporter，可修改为上送到ZipKin的Reporter。服务名，传播模式等都可以通过Builder定制创建。
public abstract class Tracing implements Closeable {
//Tracing 内部类 Builder见下方，许多组件都给了缺省定义
public static final class Builder {
//服务名与服务器Ip
String localServiceName = "unknown", localIp;
int localPort; // 服务端口
// reporter，用于处理（常见上报给ZipKin或打印到本地）链路信息
Reporter<zipkin2.Span> spanReporter;
Clock clock;//用于计时
//采样器，用于定义采样规则，默认全样采集
Sampler sampler = Sampler.ALWAYS_SAMPLE;
//用于获取当前 TraceContext ，默认使用了 InheritableThreadLocal，支持复制到异步线程
CurrentTraceContext currentTraceContext = CurrentTraceContext.Default.inheritable();
//顾名思义，traceId是否128bit，是否支持Join一个跨度
boolean traceId128Bit = false, supportsJoin = true;
//传播工厂，用于定义传播规则，如何注入与提取等。
Propagation.Factory propagationFactory = B3Propagation.FACTORY;
//错误处理器
ErrorParser errorParser = new ErrorParser();
//span结束回调器
List<FinishedSpanHandler> finishedSpanHandlers = new ArrayList<>();
//执行build方法创建 Tracing
public Tracing build() {
//根据不同的jdk获取clock（Brave支持Jdk1.6+）
if (clock == null) clock = Platform.get().clock();
//根据不同的jdk获取ip
if (localIp == null) localIp = Platform.get().linkLocalIp();
//默认reporter就是在此处定义了
if (spanReporter == null) spanReporter = new LoggingReporter();
//将 Builder 传入创建 Tracing，Default 是 Tracing一个内部类，见下方代码
return new Default(this);
}
}
static final class Default extends Tracing {
//代码太长，见下方代码块，单独说Default
}
}
static final class Default extends Tracing {
final Tracer tracer;//Tracer 可以理解为链路对象，用于操作span
final Propagation.Factory propagationFactory;
final Propagation<String> stringPropagation;
final CurrentTraceContext currentTraceContext;
final Sampler sampler;
final Clock clock;
final ErrorParser errorParser;
final AtomicBoolean noop;
Default(Builder builder) {
//初始化过程先默认从builder中获取所需对象
this.clock = builder.clock;
this.errorParser = builder.errorParser;
this.propagationFactory = builder.propagationFactory;
this.stringPropagation = builder.propagationFactory.create(Propagation.KeyFactory.STRING);
this.currentTraceContext = builder.currentTraceContext;
this.sampler = builder.sampler;
this.noop = new AtomicBoolean();
List<FinishedSpanHandler> finishedSpanHandlers = builder.finishedSpanHandlers;
// If a Zipkin reporter is present, it is invoked after the user-supplied finished span handlers.
FinishedSpanHandler zipkinFirehose = FinishedSpanHandler.NOOP;
if (builder.spanReporter != Reporter.NOOP) {//若 reporter 不是空操作
zipkinFirehose = new ZipkinFinishedSpanHandler(builder.spanReporter, errorParser,
builder.localServiceName, builder.localIp, builder.localPort);
finishedSpanHandlers = new ArrayList<>(finishedSpanHandlers);
finishedSpanHandlers.add(zipkinFirehose);
}
// 将所有的 FinishedSpanHandler 归集成一个 finishedSpanHandler
FinishedSpanHandler finishedSpanHandler =
FinishedSpanHandlers.noopAware(FinishedSpanHandlers.compose(finishedSpanHandlers), noop);
//创建一个 Tracer，差不多也是将各种组件初始化到 Tracer 中
this.tracer = new Tracer(
builder.clock,
builder.propagationFactory,
finishedSpanHandler,
new PendingSpans(clock, zipkinFirehose, noop),
builder.sampler,
builder.currentTraceContext,
builder.traceId128Bit || propagationFactory.requires128BitTraceId(),
builder.supportsJoin && propagationFactory.supportsJoin(),
finishedSpanHandler.alwaysSampleLocal(),
noop
);
maybeSetCurrent();//确保Tracing 唯一
}
private void maybeSetCurrent() {
if (current != null) return;
synchronized (Tracing.class) {
if (current == null) current = this;
}
}
}
Tracing 的解读基本完成，看Tracing的定义，基本清楚知道整个链路跟踪流程中会使用到的组件有哪些，他们的作用也大概能得知。
下一章继续讲 Tracer。
网友评论