opentelemetry-java-instrumentation/docs/contributing/using-instrumenter-api.md

Using the Instrumenter API

The Instrumenter encapsulates the entire logic for gathering telemetry, from collecting the data, to starting and ending spans, to recording values using metrics instruments. The Instrumenter public API contains only three methods: shouldStart(), start() and end(). The class is designed to decorate the actual invocation of the instrumented library code; shouldStart() and start() methods are to be called at the start of the request processing, while end() must be called when processing ends and a response arrives, or when it fails with an error. The Instrumenter is a generic class parameterized with REQUEST and RESPONSE types. They represent the input and output of the instrumented operation. Instrumenter can be configured with various extractors that can enhance or modify the telemetry data.

Check if any telemetry should be generated for the operation using shouldStart()

The first method, which needs to be called before any other Instrumenter method, is shouldStart(). It determines whether the operation should be instrumented for telemetry or not. The Instrumenter framework implements several suppression rules that prevent generating duplicate telemetry; for example the same HTTP server request always produces a single HTTP SERVER span.

The shouldStart() method accepts the current OpenTelemetry Context and the instrumented library REQUEST type. Consider the following example:

Response decoratedMethod(Request request) {
  Context parentContext = Context.current();
  if (!instrumenter.shouldStart(parentContext, request)) {
    return actualMethod(request);
  }

  // ...
}

If the shouldStart() method returns false, none of the remaining Instrumenter methods should be called.

Start an instrumented operation using start()

When shouldStart() returns true, you can use start() to initiate the instrumented operation. The start() method begins gathering telemetry about the instrumented library function that's being invoked. It starts the Span and begins recording the metrics (if any are registered in the used Instrumenter instance).

The start() method accepts the current OpenTelemetry Context and the instrumented library REQUEST type, and returns the new OpenTelemetry Context that should be made current until the instrumented operation ends. Consider the following example:

Response decoratedMethod(Request request) {
  // ...

  Context context = instrumenter.start(parentContext, request);
  try (Scope scope = context.makeCurrent()) {
    // ...
  }
}

The newly started context is made current, and inside its scope the actual library method is called.

End an instrumented operation using end()

The end() method is called when the instrumented operation finished. It is of extreme importance for this method to be always called after start(). Calling start() without later end() will result in inaccurate or wrong telemetry and context leaks. The end() method ends the current span and finishes recording the metrics (if any are registered in the Instrumenter instance).

The end() method accepts several arguments:

• The OpenTelemetry Context that was returned by the start() method.
• The REQUEST instance that started the processing.
• Optionally, the RESPONSE instance that ends the processing - it may be null in case it was not received or an error has occurred.
• Optionally, a Throwable error that was thrown by the operation.

Consider the following example:

Response decoratedMethod(Request request) {
  Context parentContext = Context.current();
    if (!instrumenter.shouldStart(parentContext, request)) {
    return actualMethod(request);
  }

  Context context = instrumenter.start(parentContext, request);
  try (Scope scope = context.makeCurrent()) {
    Response response = actualMethod(request);
    instrumenter.end(context, request, response, null);
    return response;
  } catch (Throwable error) {
    instrumenter.end(context, request, null, error);
    throw error;
  }
}

In the code sample the context returned by the start() method is passed along to the end() method. The end() method is always called, regardless of the outcome of the decorated actualMethod(), be it a valid response or an error.

Constructing a new Instrumenter using an InstrumenterBuilder

An Instrumenter can be obtained by calling its static builder() method and using the returned InstrumenterBuilder to configure captured telemetry and apply customizations. The builder() method accepts three arguments:

• An OpenTelemetry instance, which is used to obtain the Tracer and Meter objects.
• The instrumentation name, which indicates the instrumentation library name, not the instrumented library name. The value passed here should uniquely identify the instrumentation library so that during troubleshooting it's possible to determine where the telemetry came from. Read more about instrumentation libraries in the OpenTelemetry specification.
• A SpanNameExtractor that determines the span name.

An Instrumenter can be built from several smaller components. The following subsections describe all interfaces that can be used to customize an Instrumenter.

Set the instrumentation version and OpenTelemetry schema URL

By setting the instrumentation library version, you let users identify which version of your instrumentation produced the telemetry. Make sure you always provide the version to the Instrumenter. You can do this in two ways:

• By calling the setInstrumentationVersion() method on the InstrumenterBuilder.

• By making sure that the JAR file with your instrumentation library contains a properties file in the META-INF/io/opentelemetry/instrumentation/ directory. You must name the file ${instrumentationName}.properties, where ${instrumentationName} is the name of the instrumentation library passed to the Instrumenter#builder() method. The file must contain a single property, version. For example:

  # META-INF/io/opentelemetry/instrumentation/my-instrumentation.properties
  version=1.2.3
  

The Instrumenter automatically detects the properties file and determines the instrumentation version based on its name.

If the Instrumenter adheres to a specific OpenTelemetry schema, you can set the schema URL using the setSchemaUrl() method on the InstrumenterBuilder. To learn more about the OpenTelemetry schemas see the Overview.

Name the spans using the SpanNameExtractor

A SpanNameExtractor is a simple functional interface that accepts the REQUEST type and returns the span name. For more detailed guidelines on span naming please take a look at the Span specification and the tracing semantic conventions.

Consider the following example:

class MySpanNameExtractor implements SpanNameExtractor<Request> {

  @Override
  public String extract(Request request) {
    return request.getOperationName();
  }
}

The example SpanNameExtractor implementation takes a fitting value provided by the request type and uses it as a span name. Notice that SpanNameExtractor is a @FunctionalInterface: instead of implementing it as a separate class you can just pass Request::getOperationName to the builder() method.

Add attributes to span and metric data with the AttributesExtractor

An AttributesExtractor is responsible for extracting span and metric attributes when the processing starts and ends. It contains two methods:

• The onStart() method is called when the instrumented operation starts. It accepts two parameters: an AttributesBuilder instance and the incoming REQUEST instance.
• The onEnd() method is called when the instrumented operation ends. It accepts the same two parameters as onStart() and also an optional RESPONSE and an optional Throwable error.

The aim of both methods is to extract interesting attributes from the received request (and response or error) and set them into the builder. In general, it is better to populate attributes onStart(), as these attributes will be available to the Sampler.

Consider the following example:

class MyAttributesExtractor implements AttributesExtractor<Request, Response> {

  private static final AttributeKey<String> NAME = stringKey("mylib.name");
  private static final AttributeKey<Long> COUNT = longKey("mylib.count");

  @Override
  public void onStart(AttributesBuilder attributes, Request request) {
    set(attributes, NAME, request.getName());
  }

  @Override
  public void onEnd(
      AttributesBuilder attributes,
      Request request,
      @Nullable Response response,
      @Nullable Throwable error) {
    if (response != null) {
      set(attributes, COUNT, response.getCount());
    }
  }
}

The sample AttributesExtractor implementation above sets two attributes: one extracted from the request, one from the response. It is recommended to keep AttributeKey instances as static final constants and reuse them. Creating a new key each time an attribute is set risks introducing unnecessary overhead.

You can add an AttributesExtractor to the InstrumenterBuilder by using the addAttributesExtractor() or addAttributesExtractors() methods.

Set span status by setting the SpanStatusExtractor

By default, the span status is set to StatusCode.ERROR when a Throwable error occurs, and to StatusCode.UNSET in all other cases. Setting a custom SpanStatusExtractor allows to customize this behavior.

The SpanStatusExtractor interface has only one method extract() that accepts the REQUEST, an optional RESPONSE and an optional Throwable error. It is supposed to return a StatusCode that will be set when the span wrapping the instrumented operation ends. Consider the following example:

class MySpanStatusExtractor implements SpanStatusExtractor<Request, Response> {

  @Override
  public StatusCode extract(
      Request request,
      @Nullable Response response,
      @Nullable Throwable error) {
    if (response != null) {
      return response.isSuccessful() ? StatusCode.OK : StatusCode.ERROR;
    }
    return SpanStatusExtractor.getDefault().extract(request, response, error);
  }
}

The sample SpanStatusExtractor implementation above returns a custom StatusCode depending on the operation result, encoded in the response class. If the response was not present (for example, because of an error) it falls back to the default behavior, represented by the SpanStatusExtractor.getDefault() method.

You can set the SpanStatusExtractor in the InstrumenterBuilder by using the setSpanStatusExtractor() method.

Add span links using the SpanLinksExtractor

The SpanLinksExtractor interface can be used to add links to other spans when the instrumented operation starts. It has a single extract() method that receives the following arguments:

• A SpanLinkBuilder that can be used to add the links.
• The parent Context that was passed in to Instrumenter#start().
• The REQUEST instance that was passed in to Instrumenter#start().

You can read more about span links and their use cases here.

Consider the following example:

class MySpanLinksExtractor implements SpanLinksExtractor<Request> {

  @Override
  public void extract(SpanLinksBuilder spanLinks, Context parentContext, Request request) {
    for (RelatedOperation op : request.getRelatedOperations()) {
      spanLinks.addLink(op.getSpanContext());
    }
  }
}

In the SpanLinksExtractor sample implementation, the request object uses a convenient getRelatedOperations() method to find all spans that should be linked to the newly created one. When such a function doesn't exist, you need to construct SpanContext by extracting information from a data structure representing request headers or metadata.

You can add a SpanLinksExtractor to the InstrumenterBuilder by using the addSpanLinksExtractor() method.

Discard wrapper exception types with the ErrorCauseExtractor

When an error occurs, the root cause might be hidden behind several "wrapper" exception types, like an ExecutionException or a CompletionException from the Java standard library. By default, the known wrapper exception types from the JDK are removed from the captured error. To remove other wrapper exceptions, like the ones provided by the instrumented library, you can implement the ErrorCauseExtractor, which has the following features:

• It has only one method extractCause() that is responsible for stripping the unnecessary exception layers and extracting the actual error that caused the operation to fail.
• It accepts a Throwable and returns a Throwable.

Consider the following example:

class MyErrorCauseExtractor implements ErrorCauseExtractor {

  @Override
  public Throwable extract(Throwable error) {
    if (error instanceof MyLibWrapperException && error.getCause() != null) {
      error = error.getCause();
    }
    return ErrorCauseExtractor.jdk().extractCause(error);
  }
}

The example ErrorCauseExtractor implementation checks whether the error is an instance of MyLibWrapperException and has a cause, in which case it unwraps it. The error.getCause() != null check is very relevant: if the extractor did not verify both conditions, it could accidentally remove the whole exception, making the instrumentation miss an error and thus radically changing the captured telemetry. Next, the extractor falls back to the default jdk() implementation that removes the known JDK wrapper exception types.

You can set the ErrorCauseExtractor in the InstrumenterBuilder using the setErrorCauseExtractor() method.

Register metrics by implementing the OperationMetrics and OperationListener

If you need to add metrics to the Instrumenter you can implement the OperationMetrics and OperationListener interfaces. OperationMetrics is simply a factory interface for the OperationListener - it receives an OpenTelemetry Meter and returns a new listener. The OperationListener contains two methods:

• onStart() that gets executed when the instrumented operation starts. It returns a Context - it can be used to store internal metrics state that should be propagated to the onEnd() call, if needed.
• onEnd() that gets executed when the instrumented operation ends.

Both methods accept a Context, an instance of Attributes that contains either attributes computed on instrumented operation start or end, and the start and end nanoseconds timestamp that can be used to accurately compute the duration.

Consider the following example:

class MyOperationMetrics implements OperationListener {

  static OperationMetrics get() {
    return MyOperationMetrics::new;
  }

  private final LongUpDownCounter activeRequests;

  MyOperationMetrics(Meter meter) {
    activeRequests = meter
        .upDownCounterBuilder("mylib.active_requests")
        .setUnit("{requests}")
        .build();
  }

  @Override
  public Context onStart(Context context, Attributes startAttributes, long startNanos) {
    activeRequests.add(1, startAttributes);
    return context.with(new MyMetricsState(startAttributes));
  }

  @Override
  public void onEnd(Context context, Attributes endAttributes, long endNanos) {
    MyMetricsState state = MyMetricsState.get(context);
    activeRequests.add(1, state.startAttributes());
  }
}

The sample class listed above implements the OperationMetrics factory interface in the static get() method. The listener implementation uses a counter to measure the number of requests that are currently in flight. Notice that the state between onStart() and onEnd() method is shared using the MyMetricsState class (a mostly trivial data class, not listed in the example above), passed between the methods using the Context.

You can add OperationMetrics to the InstrumenterBuilder using the addOperationMetrics() method.

Enrich the operation Context with the ContextCustomizer

In some rare cases, there is a need to enrich the Context before it is returned from the Instrumenter#start() method. The ContextCustomizer interface can be used to achieve that. It exposes a single onStart() method that accepts a Context, a REQUEST and Attributes extracted on the operation start, and returns a modified Context.

Consider the following example:

class MyContextCustomizer implements ContextCustomizer<Request> {

  @Override
  public Context onStart(Context context, Request request, Attributes startAttributes) {
    return context.with(new InProcessingAttributesHolder());
  }
}

The sample ContextCustomizer listed above inserts an additional InProcessingAttributesHolder to the Context before it is returned from the Instrumenter#start() method. The InProcessingAttributesHolder class, as its name implies, may be used to keep track of attributes that are not available on request start or end - for example, if the instrumented library exposes important information only during the processing. The holder class can be looked up from the current Context and filled in with that information between the instrumented operation start and end. It can be later passed as RESPONSE type (or a part of it) to the Instrumenter#end() method so that the configured AttributesExtractor can process the collected information and turn it into telemetry attributes.

You can add a ContextCustomizer to the InstrumenterBuilder using the addContextCustomizer() method.

Disable the instrumentation

In some rare cases it may be useful to completely disable the constructed Instrumenter, for example, based on a configuration property. The InstrumenterBuilder exposes a setEnabled() method for that: passing false will turn the newly created Instrumenter into a no-op instance.

Finally, set the span kind with the SpanKindExtractor and get a new Instrumenter

The Instrumenter creation process ends with calling one of the following InstrumenterBuilder methods:

• buildInstrumenter(): the returned Instrumenter will always start spans with kind INTERNAL.
• buildInstrumenter(SpanKindExtractor): the returned Instrumenter will always start spans with kind determined by the passed SpanKindExtractor.
• buildClientInstrumenter(TextMapSetter): the returned Instrumenter will always start CLIENT spans and will propagate operation context into the outgoing request.
• buildServerInstrumenter(TextMapGetter): the returned Instrumenter will always start SERVER spans and will extract the parent span context from the incoming request.
• buildProducerInstrumenter(TextMapSetter): the returned Instrumenter will always start PRODUCER spans and will propagate operation context into the outgoing request.
• buildConsumerInstrumenter(TextMapGetter): the returned Instrumenter will always start CONSUMER spans and will extract the parent span context from the incoming request.

The last four variants that create non-INTERNAL spans accept either TextMapSetter or TextMapGetter implementations as parameters. These are needed to correctly implement the context propagation between services. If you want to learn how to use and implement these interfaces, read the OpenTelemetry Java docs.

The SpanKindExtractor interface, accepted by the second variant from the list above, is a simple interface that accepts a REQUEST and returns a SpanKind that should be used when starting the span for this operation. Consider the following example:

class MySpanKindExtractor implements SpanKindExtractor<Request> {

  @Override
  public SpanKind extract(Request request) {
    return request.shouldSynchronouslyWaitForResponse() ? SpanKind.CLIENT : SpanKind.PRODUCER;
  }
}

The example SpanKindExtractor above decides whether to use PRODUCER or CLIENT based on how the request is going to be processed. This example reflects a real-life scenario: you might find similar code in a messaging library instrumentation, since according to the OpenTelemetry messaging semantic conventions the span kind should be set to CLIENT if sending the message is completely synchronous and waits for the response.