This topic describes the background, scenarios, and core features of Tracing Analysis in Function Compute. This topic also shows you how to configure the sampling rule and enable Tracing Analysis.

Background information

Alibaba Cloud Tracing Analysis adopts the OpenTracing standard and is fully compatible with open source communities. Tracing Analysis provides developers of distributed applications with various features. For example, it enables developers to query and diagnose distributed traces, dynamically discover distributed topologies, and summarize application performance in real time.

Function Compute integrates with Tracing Analysis. You can use Jaeger to upload trace information so that you can track function invocation. Tracing Analysis helps you analyze and diagnose performance bottlenecks in a serverless architecture. This improves development and diagnostics efficiency in serverless scenarios.


After you configure Tracing Analysis for a service in Function Compute, you can record the execution time of a request in Function Compute. You can also view the cold start time for a function and record the execution time of a function. Tracing Analysis helps you troubleshoot the following issues:

  • Identify the performance bottlenecks of a function if each execution of the function times out.
  • Identify the cause of the long end-to-end latency if the execution time of a function is short.
  • Analyze and diagnose the performance bottlenecks of a function if a request involves multiple cloud services in a distributed system.

Core features

Tracing Analysis in Function Compute connects the entire trace. Tracing Analysis provides the following core features:

  • Automatically records the time that is spent on key steps in Function Compute.
  • Connects to an upstream service. If a request header contains a SpanContext, Function Compute creates a child span based on the SpanContext.
  • Connects to a downstream service. Function Compute imports the SpanContext to the function context to help you track the internal trace of a function.
  • Allows you to view application topologies.
  • Allows you to view the execution of an invalid API operation and identify the cause of the error.

Automatically record the time that is spent on key steps in Function Compute

After Tracing Analysis is enabled for a service, you can view the following trace by default.


The trace contains the following spans:

  • InvokeFunction: the total execution time of the current request in Function Compute.
  • ColdStart: the cold start time of the system for the function. A cold start does not occur every time you invoke a function, and occurs only when you reapply for the execution environment.
    • PrepareCode: the time that is taken to download code or a custom image for the function. If the time indicated by the PrepareCode span is too long, simplify the code package to reduce the time that is taken to prepare code.
    • RuntimeInitialization: the time that is taken to start the execution environment, including the time that is taken to start an instance and the time that is taken to perform health check on the instance. For a custom runtime and a custom image, if the time indicated by the RuntimeInitialization span is too long, check the startup behavior of the corresponding HTTP server and image.
  • Initializer: the time that is taken to execute the initializer function. The initializer function is executed only when the container is cold started.
  • Invocation: the time that is taken to execute the function. You can check the context of Invocation in the function to obtain the execution time.

The function generated by Function Compute is named in the format of FC:ServiceName/FunctionName.

If a request does not encounter a cold start, the trace does not contain the ColdStart and Initializer spans. The following figure shows the trace.


Connect to an upstream service

After Tracing Analysis is enabled for a service, if a request header contains a SpanContext, Function Compute creates a child span based on the SpanContext.

Function Compute identifies the following SpanContext headers:

  • x-fc-tracing-opentracing-span-context: transmits the SpanContext.
  • x-fc-tracing-opentracing-span-context-baggage-: transmits a cross-context baggage.

    If multiple baggages are to be transmitted, you must transmit multiple headers, such as x-fc-tracing-opentracing-span-context-baggage-key1: val1 and tracing-opentracing-span-context-baggage-key2: val2.

The SpanContext can be injected by adding SpanContext headers when the function is invoked.

Node.js is used in the following example:

'use strict';
const FCClient = require('@alicloud/fc2');

var client = new FCClient('<account id>', {
    accessKeyID: '<access key id>',
    accessKeySecret: '<access key secret>',
    region: 'cn-shanghai',

var serviceName = '<service name>';
var funcName = '<function name>';

async function test() {
    try {

        // inject spanContext headers
        var headers = {
            'x-fc-tracing-opentracing-span-context': '124ed43254b54966:124ed43254b5****:0:1',
            'x-fc-tracing-opentracing-span-context-baggage-key': 'val'
        var resp = await client.invokeFunction(serviceName, functionName, 'event', headers = headers);
    } catch (err) {

Connect to a downstream service

Function Compute imports a SpanContext to a function to help you track the internal trace of the function.

  • In a built-in runtime, the SpanContext can be obtained by using context.tracing.
  • In a custom runtime or custom container, the SpanContext in Function Compute can be obtained from the request header.
    • x-fc-tracing-opentracing-span-context: the SpanContext of InvokeFunction in Function Compute. A span is created based on the SpanContext in the function.
    • x-fc-tracing-opentracing-span-baggages: the Base64-encoded cross-context baggages.
    • x-fc-tracing-jaeger-endpoint: the endpoint of the Jaeger server. You can directly upload the span in the function to this endpoint.

For example, context.tracing has the following structure:

  "openTracingSpanContext": "5f22f355044a957a:5708f3a95a4ed10:5f22f355044a****:1",
  "openTracingSpanBaggages": {
    "key1": "val1",
    "key2": "val2"
  "jaegerEndpoint": ""

The following code provides an example on how to obtain the SpanContext from the function:

  • Node.js:
    exports.handler = (event, context, callback) => {
      var params = {
            openTracingSpanContext: context.tracing.openTracingSpanContext,
            // jaegerEndpoint is confidential, do not print it out easily
            // jaegerEndpoint:context.tracing.jaegerEndpoint
      console.log('tracing params',params)
  • PHP:
    function handler($event, $ctx) {
        $logger = $GLOBALS['fcLogger'];
        $openTracingSpanContext = $ctx['tracing']['openTracingSpanContext'];
        $openTracingSpanBaggages = $ctx['tracing']['openTracingSpanBaggages'];
        // jaegerEndpoint is confidential, do not print it out easily
        $jaegerEndpoint = $ctx['tracing']['jaegerEndpoint'];
      return 'success';
  • Java:
    package example;
    import com.aliyun.fc.runtime.Context;
    import com.aliyun.fc.runtime.StreamRequestHandler;
    public class HelloFC implements StreamRequestHandler {
        public void handleRequest(
                InputStream inputStream, OutputStream outputStream, Context context) throws IOException {
            String spanContext = context.getTracing().getOpenTracingSpanContext();
            String jaegerEndpoint = context.getTracing().getJaegerEndpoint();
            String spanBaggage = context.getTracing().getOpenTracingSpanBaggages().get("key1");
            outputStream.write(new String("success").getBytes());
  • In a custom runtime or custom container, you need only to obtain the header. Go is used in the following example:
      spanContext := req.Header.Get('x-fc-tracing-opentracing-span-context')
      spanBaggages := req.Header.Get('x-fc-tracing-opentracing-span-baggages')
      jaegerEndpoint := req.Header.Get('x-fc-tracing-jaeger-endpoint')

Customize the sampling rule

The service name that corresponds to Function Compute is fc-tracing. By default, RatelimitingSampler is used, and sampling is performed at a rate of one request per second.

To customize the sampling rule, log on to the Tracing Analysis console and configure the remote sampling rule. For more information, see Use Jaeger to configure remote sampling policies. After the configuration is complete, Function Compute uses the remote sampling rule that you configured to perform sampling.

By default, Function Compute uses the following sampling rule:

  "default_strategy": {
    "type": "ratelimiting",
    "param": 1,

Enabling method

For information about how to enable Tracing Analysis, see Configure Tracing Analysis.