This guide will walk you through instrumenting a Python application with ``opentelemetry-python``.

For more elaborate examples, see `examples`.

Hello world: emitting a trace to your console

The SDK provides an implementation of those interfaces, designed to be generic and extensible enough

that in many situations, the SDK will be sufficient.

Once installed, we can now utilize the packages to emit spans from your application. A span

or a database call. Once instrumented, the application owner can extract helpful information such as

how long the action took, or add arbitrary attributes to the span that may provide more insight for debugging.

Here's an example of a script that emits a trace containing three named spans: "foo", "bar", and "baz":

We can run it, and see the traces print to your console:

While a given span is active, newly-created spans will inherit the active span's trace ID, options, and other attributes of its context.

A span without a parent is called the "root span", and a trace is comprised of one root span and its descendants.

In the example above, the OpenTelemetry Python library creates one trace containing three spans and prints it to STDOUT.

The example above does emit information about all spans, but the output is a bit hard to read.

In common cases, you would instead *export* this data to an application performance monitoring backend, to be visualized and queried.

It is also common to aggregate span and trace information from multiple services into a single database, so that actions that require multiple services can still all be visualized together.

This concept is known as distributed tracing. One such distributed tracing backend is known as Jaeger.

The Jaeger project provides an all-in-one docker container that provides a UI, database, and consumer. Let's bring

it up now:

This will start Jaeger on port 16686 locally, and expose Jaeger thrift agent on port 6831. You can visit it at http://localhost:16686.

With this backend up, your application will now need to export traces to this system. ``opentelemetry-sdk`` does not provide an exporter

for Jaeger, but you can install that as a separate package:

Once installed, update your code to import the Jaeger exporter, and use that instead:

Run the script:

You can then visit the jaeger UI, see you service under "services", and find your traces!

Integrations example with Flask

The above is a great example, but it's very manual. Within the telemetry space, there are common actions that one wants to instrument:

To help instrument common scenarios, opentelemetry also has the concept of "instrumentations": packages that are designed to interface

with a specific framework or library, such as Flask and psycopg2. A list of the currently curated extension packages can be found `at the Contrib repo <https://github.com/open-telemetry/opentelemetry-python-contrib/tree/master/instrumentation>`_.

We will now instrument a basic Flask application that uses the requests library to send HTTP requests. First, install the instrumentation packages themselves:

And let's write a small Flask application that sends an HTTP request, activating each instrumentation during the initialization:

Now run the above script, hit the root url (http://localhost:5000/) a few times, and watch your spans be emitted!

Configure Your HTTP Propagator (b3, Baggage)

To enable this, OpenTelemetry has the concept of `propagators <https://github.com/open-telemetry/opentelemetry-specification/blob/master/specification/context/api-propagators.md>`_,

which provide a common method to encode and decode span information from a request and response,

respectively.

By default, opentelemetry-python is configured to use the `W3C Trace Context <https://www.w3.org/TR/trace-context/>`_

HTTP headers for HTTP requests. This can be configured to leverage different propagators. Here's

what about a more aggregated perspective? OpenTelemetry provides supports for metrics, a time series

of numbers that might express things such as CPU utilization, request count for an HTTP server, or a

All metrics can be annotated with labels: additional qualifiers that help describe what

The following is an example of emitting metrics to console, in a similar fashion to the trace example:

The sleeps will cause the script to take a while, but running it should yield:

Using Prometheus

Similar to traces, it is really valuable for metrics to have its own data store to help visualize and query the data. A common solution for this is

`Prometheus <https://prometheus.io/>`_.

Let's start by bringing up a Prometheus instance ourselves, to scrape our application. Write the following configuration:

And start a docker container for it:

For our Python application, we will need to install an exporter specific to Prometheus:

And use that instead of the `ConsoleMetricsExporter`:

The Prometheus server will run locally on port 8000, and the instrumented code will make metrics available to Prometheus via the `PrometheusMetricsExporter`.

Using the OpenTelemetry Collector for traces and metrics

Although it's possible to directly export your telemetry data to specific backends, you may more complex use cases, including:

* having a single telemetry sink shared by multiple services, to reduce overhead of switching exporters

* aggregating metrics or traces across multiple services, running on multiple hosts

To see how this works in practice, let's start the Collector locally. Write the following file:

Start the docker container:

And execute the following script: