Manual Instrumentation

Libraries that want to export telemetry data using OpenTelemetry MUST only depend on the opentelemetry-api package and should never configure or depend on the OpenTelemetry SDK. The SDK configuration must be provided by Applications which should also depend on the opentelemetry-sdk package, or any other implementation of the OpenTelemetry API. This way, libraries will obtain a real implementation only if the user application is configured for it. For more details, check out the Library Guidelines.

Set up

The first step is to get a handle to an instance of the OpenTelemetry interface.

If you are an application developer, you need to configure an instance of the OpenTelemetrySdk as early as possible in your application. This can be done using the OpenTelemetrySdk.builder() method.

For example:

SdkTracerProvider sdkTracerProvider = SdkTracerProvider.builder()

SdkMeterProvider meterProvider = SdkMeterProvider.builder()

OpenTelemetry openTelemetry = OpenTelemetrySdk.builder()

As an aside, if you are writing library instrumentation, it is strongly recommended that you provide your users the ability to inject an instance of OpenTelemetry into your instrumentation code. If this is not possible for some reason, you can fall back to using an instance from the GlobalOpenTelemetry class. Note that you can’t force end-users to configure the global, so this is the most brittle option for library instrumentation.


In the following, we present how to trace code using the OpenTelemetry API.

Note: Methods of the OpenTelemetry SDK should never be called.

First, a Tracer must be acquired, which is responsible for creating spans and interacting with the Context. A tracer is acquired by using the OpenTelemetry API specifying the name and version of the library instrumenting the instrumented library or application to be monitored. More information is available in the specification chapter Obtaining a Tracer.

Tracer tracer =
    openTelemetry.getTracer("instrumentation-library-name", "1.0.0");

Important: the “name” and optional version of the tracer are purely informational. All Tracers that are created by a single OpenTelemetry instance will interoperate, regardless of name.

Create a basic Span

To create a basic span, you only need to specify the name of the span. The start and end time of the span is automatically set by the OpenTelemetry SDK.

Span span = tracer.spanBuilder("my span").startSpan();
// put the span into the current Context
try {
  // do something...
} finally {

Create nested Spans

Most of the time, we want to correlate spans for nested operations. OpenTelemetry supports tracing within processes and across remote processes. For more details how to share context between remote processes, see Context Propagation.

For a method a calling a method b, the spans could be manually linked in the following way:

void parentOne() {
  Span parentSpan = tracer.spanBuilder("parent").startSpan();
  try {
  } finally {

void childOne(Span parentSpan) {
  Span childSpan = tracer.spanBuilder("child")
  try {
    // do stuff
  } finally {

The OpenTelemetry API offers also an automated way to propagate the parent span on the current thread:

void parentTwo() {
  Span parentSpan = tracer.spanBuilder("parent").startSpan();
  try(Scope scope = parentSpan.makeCurrent()) {
  } finally {
void childTwo() {
  Span childSpan = tracer.spanBuilder("child")
    // NOTE: setParent(...) is not required;
    // `Span.current()` is automatically added as the parent
  try(Scope scope = childSpan.makeCurrent()) {
    // do stuff
  } finally {

To link spans from remote processes, it is sufficient to set the Remote Context as parent.

Span childRemoteParent = tracer.spanBuilder("Child").setParent(remoteContext).startSpan();

Get the current span

Sometimes it’s helpful to do something with the current/active span at a particular point in program execution.

Span span = Span.current()

And if you want the current span for a particular Context object:

Span span = Span.fromContext(context)

Span Attributes

In OpenTelemetry spans can be created freely and it’s up to the implementor to annotate them with attributes specific to the represented operation. Attributes provide additional context on a span about the specific operation it tracks, such as results or operation properties.

Span span = tracer.spanBuilder("/resource/path").setSpanKind(SpanKind.CLIENT).startSpan();
span.setAttribute("http.method", "GET");
span.setAttribute("http.url", url.toString());

Some of these operations represent calls that use well-known protocols like HTTP or database calls. For these, OpenTelemetry requires specific attributes to be set. The full attribute list is available in the Semantic Conventions in the cross-language specification.

Create Spans with events

Spans can be annotated with named events that can carry zero or more Span Attributes, each of which is itself a key:value map paired automatically with a timestamp.

Attributes eventAttributes = Attributes.of(
    AttributeKey.stringKey("key"), "value",
    AttributeKey.longKey("result"), 0L);

span.addEvent("End Computation", eventAttributes);

A Span may be linked to zero or more other Spans that are causally related. Links can be used to represent batched operations where a Span was initiated by multiple initiating Spans, each representing a single incoming item being processed in the batch.

Span child = tracer.spanBuilder("childWithLink")

For more details how to read context from remote processes, see Context Propagation.

Set span status

A status can be set on a span, typically used to specify that a span has not completed successfully - SpanStatus.Error. In rare scenarios, you could override the Error status with OK, but don’t set OK on successfully-completed spans.

The status can be set at any time before the span is finished:

Span span = tracer.spanBuilder("my span").startSpan();
// put the span into the current Context
try (Scope scope = span.makeCurrent()) {
	// do something
} catch (Throwable t) {
  span.setStatus(StatusCode.ERROR, "Something bad happened!");
  throw t;
} finally {
  span.end(); // Cannot set a span after this call

Record exceptions in spans

It can be a good idea to record exceptions when they happen. It’s recommended to do this in conjunction with setting span status.

Span span = tracer.spanBuilder("my span").startSpan();
// put the span into the current Context
try (Scope scope = span.makeCurrent()) {
	// do something
} catch (Throwable throwable) {
  span.setStatus(StatusCode.ERROR, "Something bad happened!");
} finally {
  span.end(); // Cannot set a span after this call

This will capture things like the current stack trace in the span.

Context Propagation

OpenTelemetry provides a text-based approach to propagate context to remote services using the W3C Trace Context HTTP headers.

The following presents an example of an outgoing HTTP request using HttpURLConnection.

// Tell OpenTelemetry to inject the context in the HTTP headers
TextMapSetter<HttpURLConnection> setter =
  new TextMapSetter<HttpURLConnection>() {
    public void set(HttpURLConnection carrier, String key, String value) {
        // Insert the context as Header
        carrier.setRequestProperty(key, value);

URL url = new URL("");
Span outGoing = tracer.spanBuilder("/resource").setSpanKind(SpanKind.CLIENT).startSpan();
try (Scope scope = outGoing.makeCurrent()) {
  // Use the Semantic Conventions.
  // (Note that to set these, Span does not *need* to be the current instance in Context or Scope.)
  outGoing.setAttribute(SemanticAttributes.HTTP_METHOD, "GET");
  outGoing.setAttribute(SemanticAttributes.HTTP_URL, url.toString());
  HttpURLConnection transportLayer = (HttpURLConnection) url.openConnection();
  // Inject the request with the *current*  Context, which contains our current Span.
  openTelemetry.getPropagators().getTextMapPropagator().inject(Context.current(), transportLayer, setter);
  // Make outgoing call
} finally {

Similarly, the text-based approach can be used to read the W3C Trace Context from incoming requests. The following presents an example of processing an incoming HTTP request using HttpExchange.

TextMapGetter<HttpExchange> getter =
  new TextMapGetter<>() {
    public String get(HttpExchange carrier, String key) {
      if (carrier.getRequestHeaders().containsKey(key)) {
        return carrier.getRequestHeaders().get(key).get(0);
      return null;

   public Iterable<String> keys(HttpExchange carrier) {
     return carrier.getRequestHeaders().keySet();
public void handle(HttpExchange httpExchange) {
  // Extract the SpanContext and other elements from the request.
  Context extractedContext = openTelemetry.getPropagators().getTextMapPropagator()
        .extract(Context.current(), httpExchange, getter);
  try (Scope scope = extractedContext.makeCurrent()) {
    // Automatically use the extracted SpanContext as parent.
    Span serverSpan = tracer.spanBuilder("GET /resource")
    try {
      // Add the attributes defined in the Semantic Conventions
      serverSpan.setAttribute(SemanticAttributes.HTTP_METHOD, "GET");
      serverSpan.setAttribute(SemanticAttributes.HTTP_SCHEME, "http");
      serverSpan.setAttribute(SemanticAttributes.HTTP_HOST, "localhost:8080");
      serverSpan.setAttribute(SemanticAttributes.HTTP_TARGET, "/resource");
      // Serve the request
    } finally {


Spans provide detailed information about your application, but produce data that is proportional to the load on the system. In contrast, metrics combine individual measurements into aggregations, and produce data which is constant as a function of system load. The aggregations lack details required to diagnose low level issues, but complement spans by helping to identify trends and providing application runtime telemetry.

The metrics API defines a variety of instruments. Instruments record measurements, which are aggregated by the metrics SDK and eventually exported out of process. Instruments come in synchronous and asynchronous varieties. Synchronous instruments record measurements as they happen. Asynchronous instrument register a callback, which is invoked once per collection, and which records measurements at that point in time. The following instruments are available:

  • LongCounter/DoubleCounter: records only positive values, with synchronous and asynchronous options. Useful for counting things, such as the number of bytes sent over a network. Counter measurements are aggregated to always-increasing monotonic sums by default.
  • LongUpDownCounter/DoubleUpDownCounter: records positive and negative values, with synchronous and asynchronous options. Useful for counting things that go up and down, like the size of a queue. Up down counter measurements are aggregated to non-monotonic sums by default.
  • LongGauge/DoubleGauge: measures an instantaneous value with an asynchronous callback. Useful for recording values that can’t be merged across attributes, like CPU utilization percentage. Gauge measurements are aggregated as gauges by default.
  • LongHistogram/DoubleHistogram: records measurements that are most useful to analyze as a histogram distribution. No asynchronous option is available. Useful for recording things like the duration of time spent by an HTTP server processing a request. Histogram measurements are aggregated to explicit bucket histograms by default.

Note The asynchronous varieties of counter and up down counter assume that the registered callback is observing the cumulative sum. For example, if you register an asynchronous counter whose callback records bytes sent over a network, it must record the cumulative sum of all bytes sent over the network, rather than trying to compute and record the difference since last call.

All metrics can be annotated with attributes: additional qualifiers that help describe what subdivision of the measurements the metric represents.

The following is an example of counter usage:

OpenTelemetry openTelemetry = // obtain instance of OpenTelemetry

// Gets or creates a named meter instance
Meter meter = openTelemetry.meterBuilder("instrumentation-library-name")

// Build counter e.g. LongCounter
LongCounter counter = meter
      .setDescription("Processed jobs")

// It is recommended that the API user keep a reference to Attributes they will record against
Attributes attributes = Attributes.of(stringKey("Key"), "SomeWork");

// Record data
counter.add(123, attributes);

The following is an example of usage of an asynchronous instrument:

// Build an asynchronous instrument, e.g. Gauge
  .setDescription("CPU Usage")
  .buildWithCallback(measurement -> {
    measurement.record(getCpuUsage(), Attributes.of(stringKey("Key"), "SomeWork"));

SDK Configuration

The configuration examples reported in this document only apply to the SDK provided by opentelemetry-sdk. Other implementation of the API might provide different configuration mechanisms.

Tracing SDK

The application has to install a span processor with an exporter and may customize the behavior of the OpenTelemetry SDK.

For example, a basic configuration instantiates the SDK tracer provider and sets to export the traces to a logging stream.

SdkTracerProvider tracerProvider = SdkTracerProvider.builder()


It is not always feasible to trace and export every user request in an application. In order to strike a balance between observability and expenses, traces can be sampled.

The OpenTelemetry SDK offers four samplers out of the box:

  • AlwaysOnSampler which samples every trace regardless of upstream sampling decisions.
  • AlwaysOffSampler which doesn’t sample any trace, regardless of upstream sampling decisions.
  • ParentBased which uses the parent span to make sampling decisions, if present.
  • TraceIdRatioBased which samples a configurable percentage of traces, and additionally samples any trace that was sampled upstream.

Additional samplers can be provided by implementing the io.opentelemetry.sdk.trace.Sampler interface.

SdkTracerProvider tracerProvider = SdkTracerProvider.builder()

Span Processor

Different Span processors are offered by OpenTelemetry. The SimpleSpanProcessor immediately forwards ended spans to the exporter, while the BatchSpanProcessor batches them and sends them in bulk. Multiple Span processors can be configured to be active at the same time using the MultiSpanProcessor.

SdkTracerProvider tracerProvider = SdkTracerProvider.builder()


Span processors are initialized with an exporter which is responsible for sending the telemetry data a particular backend. OpenTelemetry offers five exporters out of the box:

  • InMemorySpanExporter: keeps the data in memory, useful for testing and debugging.
  • Jaeger Exporter: prepares and sends the collected telemetry data to a Jaeger backend via gRPC. Varieties include JaegerGrpcSpanExporter and JaegerThriftSpanExporter.
  • ZipkinSpanExporter: prepares and sends the collected telemetry data to a Zipkin backend via the Zipkin APIs.
  • Logging Exporter: saves the telemetry data into log streams. Varieties include LoggingSpanExporter and OtlpJsonLoggingSpanExporter.
  • OpenTelemetry Protocol Exporter: sends the data in OTLP to the OpenTelemetry Collector or other OTLP receivers. Varieties include OtlpGrpcSpanExporter and OtlpHttpSpanExporter.

Other exporters can be found in the OpenTelemetry Registry.

ManagedChannel jaegerChannel = ManagedChannelBuilder.forAddress("localhost", 3336)

JaegerGrpcSpanExporter jaegerExporter = JaegerGrpcSpanExporter.builder()
  .setTimeout(30, TimeUnit.SECONDS)

SdkTracerProvider tracerProvider = SdkTracerProvider.builder()

Metrics SDK

The application has to install a metric reader with an exporter, and may further customize the behavior of the OpenTelemetry SDK.

For example, a basic configuration instantiates the SDK meter provider and sets to export the metrics to a logging stream.

SdkMeterProvider meterProvider = SdkMeterProvider.builder()

Metric Reader

Metric readers read aggregated metrics.

SdkMeterProvider meterProvider = SdkMeterProvider.builder()

OpenTelemetry provides a variety of metric readers out of the box:

  • PeriodicMetricReader: reads metrics on a configurable interval and pushes to a MetricExporter.
  • InMemoryMetricReader: reads metrics into memory, useful for debugging and testing.
  • PrometheusHttpServer (alpha): an HTTP server that reads metrics and serializes to Prometheus text format.

Custom metric reader implementations are not currently supported.


The PeriodicMetricReader is paired with a metric exporter, which is responsible for sending the telemetry data to a particular backend. OpenTelemetry provides the following exporters out of the box:

  • InMemoryMetricExporter: keeps the data in memory, useful for testing and debugging.
  • Logging Exporter: saves the telemetry data into log streams. Varieties include LoggingMetricExporter and OtlpJsonLoggingMetricExporter.
  • OpenTelemetry Protocol Exporter: sends the data in OTLP to the OpenTelemetry Collector or other OTLP receivers. Varieties include OtlpGrpcMetricExporter and OtlpHttpMetricExporter.

Other exporters can be found in the OpenTelemetry Registry.


Views provide a mechanism for controlling how measurements are aggregated into metrics. They consist of an InstrumentSelector and a View. The instrument selector consists of a series of options for selecting which instruments the view applies to. Instruments can be selected by a combination of name, type, meter name, meter version, and meter schema url. The view describes how measurement should be aggregated. The view can change the name, description, the aggregation, and define the set of attribute keys that should be retained.

SdkMeterProvider meterProvider = SdkMeterProvider.builder()
      .setName("my-counter") // Select instrument(s) called "my-counter"
      .setName("new-counter-name") // Change the name to "new-counter-name"

Every instrument has a default view, which retains the original name, description, and attributes, and has a default aggregation that is based on the type of instrument. When a registered view matches an instrument, the default view is replaced by the registered view. Additional registered views that match the instrument are additive, and result in multiple exported metrics for the instrument.

Auto Configuration

To configure the OpenTelemetry SDK based on the standard set of environment variables and system properties, you can use the opentelemetry-sdk-extension-autoconfigure module.

OpenTelemetrySdk sdk = OpenTelemetrySdkAutoConfiguration.initialize();

See the supported configuration options in the module’s README.

Logging and Error Handling

OpenTelemetry uses java.util.logging to log information about OpenTelemetry, including errors and warnings about misconfigurations or failures exporting data.

By default, log messages are handled by the root handler in your application. If you have not installed a custom root handler for your application, logs of level INFO or higher are sent to the console by default.

You may want to change the behavior of the logger for OpenTelemetry. For example, you can reduce the logging level to output additional information when debugging, increase the level for a particular class to ignore errors coming from that class, or install a custom handler or filter to run custom code whenever OpenTelemetry logs a particular message.


## Turn off all OpenTelemetry logging
io.opentelemetry.level = OFF
## Turn off logging for just the BatchSpanProcessor
io.opentelemetry.sdk.trace.export.BatchSpanProcessor.level = OFF
## Log "FINE" messages for help in debugging
io.opentelemetry.level = FINE

## Sets the default ConsoleHandler's logger's level
## Note this impacts the logging outside of OpenTelemetry as well
java.util.logging.ConsoleHandler.level = FINE

For more fine-grained control and special case handling, custom handlers and filters can be specified with code.

// Custom filter which does not log errors that come from the export
public class IgnoreExportErrorsFilter implements Filter {

 public boolean isLoggable(LogRecord record) {
    return !record.getMessage().contains("Exception thrown by the export");
## Registering the custom filter on the BatchSpanProcessor
io.opentelemetry.sdk.trace.export.BatchSpanProcessor = io.opentelemetry.extension.logging.IgnoreExportErrorsFilter