Injecting Auto-instrumentation

An implementation of auto-instrumentation using the OpenTelemetry Operator.

The OpenTelemetry Operator supports injecting and configuring auto-instrumentation libraries for .NET, Java, Node.js, Python, and Go services.

Installation

First, install the OpenTelemetry Operator into your cluster.

You can do this with the Operator release manifest, the Operator helm chart, or with Operator Hub.

In most cases, you will need to install cert-manager. If you use the helm chart, there is an option to generate a self-signed cert instead.

If you want to use Go auto-instrumentation, you need to enable the feature gate. See Controlling Instrumentation Capabilities for details.

Create an OpenTelemetry Collector (Optional)

It is a best practice to send telemetry from containers to an OpenTelemetry Collector instead of directly to a backend. The Collector helps simplify secret management, decouples data export problems (such as a need to do retries) from your apps, and lets you add additional data to your telemetry, such as with the k8sattributesprocessor component. If you chose not to use a Collector, you can skip to the next section.

The Operator provides a Custom Resource Definition (CRD) for the OpenTelemetry Collector which is used to create an instance of the Collector that the Operator manages. The following example deploys the Collector as a deployment (the default), but there are other deployment modes that can be used.

When using the Deployment mode the operator will also create a Service that can be used to interact with the Collector. The name of the service is the name of the OpenTelemetryCollector resource prepended to -collector. For our example that will be demo-collector.

kubectl apply -f - <<EOF
apiVersion: opentelemetry.io/v1alpha1
kind: OpenTelemetryCollector
metadata:
  name: demo
spec:
  config: |
    receivers:
      otlp:
        protocols:
          grpc:
            endpoint: 0.0.0.0:4317
          http:
            endpoint: 0.0.0.0:4318
    processors:
      memory_limiter:
        check_interval: 1s
        limit_percentage: 75
        spike_limit_percentage: 15
      batch:
        send_batch_size: 10000
        timeout: 10s

    exporters:
      # NOTE: Prior to v0.86.0 use `logging` instead of `debug`.
      debug:

    service:
      pipelines:
        traces:
          receivers: [otlp]
          processors: [memory_limiter, batch]
          exporters: [debug]
        metrics:
          receivers: [otlp]
          processors: [memory_limiter, batch]
          exporters: [debug]
        logs:
          receivers: [otlp]
          processors: [memory_limiter, batch]
          exporters: [debug]
EOF

The above command results in a deployment of the Collector that you can use as an endpoint for auto-instrumentation in your pods.

Configure Automatic Instrumentation

To be able to manage automatic instrumentation, the Operator needs to be configured to know what pods to instrument and which automatic instrumentation to use for those pods. This is done via the Instrumentation CRD.

Creating the Instrumentation resource correctly is paramount to getting auto-instrumentation working. Making sure all endpoints and env vars are correct is required for auto-instrumentation to work properly.

.NET

The following command will create a basic Instrumentation resource that is configured specifically for instrumenting .NET services.

kubectl apply -f - <<EOF
apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
metadata:
  name: demo-instrumentation
spec:
  exporter:
    endpoint: http://demo-collector:4318
  propagators:
    - tracecontext
    - baggage
  sampler:
    type: parentbased_traceidratio
    argument: "1"
EOF

By default, the Instrumentation resource that auto-instruments .NET services uses otlp with the http/protobuf protocol. This means that the configured endpoint must be able to receive OTLP over http/protobuf. Therefore, the example uses http://demo-collector:4318, which will connect to the http port of the otlpreceiver of the Collector created in the previous step.

Excluding auto-instrumentation

By default, the .NET auto-instrumentation ships with many instrumentation libraries. This makes instrumentation easy, but could result in too much or unwanted data. If there are any libraries you do not want to use you can set the OTEL_DOTNET_AUTO_[SIGNAL]_[NAME]_INSTRUMENTATION_ENABLED=false where [SIGNAL] is the type of the signal and [NAME] is the case-sensitive name of the library.

apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
metadata:
  name: demo-instrumentation
spec:
  exporter:
    endpoint: http://demo-collector:4318
  propagators:
    - tracecontext
    - baggage
  sampler:
    type: parentbased_traceidratio
    argument: '1'
  dotnet:
    env:
      - name: OTEL_DOTNET_AUTO_TRACES_GRPCNETCLIENT_INSTRUMENTATION_ENABLED
        value: false
      - name: OTEL_DOTNET_AUTO_METRICS_PROCESS_INSTRUMENTATION_ENABLED
        value: false

Learn more

For more details, see .NET Auto Instrumentation docs.

Java

The following command creates a basic Instrumentation resource that is configured for instrumenting Java services.

kubectl apply -f - <<EOF
apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
metadata:
  name: demo-instrumentation
spec:
  exporter:
    endpoint: http://demo-collector:4318
  propagators:
    - tracecontext
    - baggage
  sampler:
    type: parentbased_traceidratio
    argument: "1"
EOF

By default, the Instrumentation resource that auto-instruments Java services uses otlp with the http/protobuf protocol. This means that the configured endpoint must be able to receive OTLP over http via protobuf payloads. Therefore, the example uses http://demo-collector:4318, which connects to the http port of the otlpreceiver of the Collector created in the previous step.

Excluding auto-instrumentation

By default, the Java auto-instrumentation ships with many instrumentation libraries. This makes instrumentation easy, but could result in too much or unwanted data. If there are any libraries you do not want to use you can set the OTEL_INSTRUMENTATION_[NAME]_ENABLED=false where [NAME] is the name of the library. If you know exactly which libraries you want to use, you can disable the default libraries by setting OTEL_INSTRUMENTATION_COMMON_DEFAULT_ENABLED=false and then use OTEL_INSTRUMENTATION_[NAME]_ENABLED=true where [NAME] is the name of the library. For more details, see Suppressing specific instrumentation.

apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
metadata:
  name: demo-instrumentation
spec:
  exporter:
    endpoint: http://demo-collector:4318
  propagators:
    - tracecontext
    - baggage
  sampler:
    type: parentbased_traceidratio
    argument: '1'
  java:
    env:
      - name: OTEL_INSTRUMENTATION_KAFKA_ENABLED
        value: false
      - name: OTEL_INSTRUMENTATION_REDISCALA_ENABLED
        value: false

Learn more

For more details, see Java agent Configuration.

Node.js

The following command creates a basic Instrumentation resource that is configured for instrumenting Node.js services.

kubectl apply -f - <<EOF
apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
metadata:
  name: demo-instrumentation
spec:
  exporter:
    endpoint: http://demo-collector:4317
  propagators:
    - tracecontext
    - baggage
  sampler:
    type: parentbased_traceidratio
    argument: "1"
EOF

By default, the Instrumentation resource that auto-instruments Node.js services uses otlp with the grpc protocol. This means that the configured endpoint must be able to receive OTLP over grpc. Therefore, the example uses http://demo-collector:4317, which connects to the grpc port of the otlpreceiver of the Collector created in the previous step.

Excluding instrumentation libraries

By default, the Node.js zero-code instrumentation has all the instrumentation libraries enabled.

To enable only specific instrumentation libraries you can use the OTEL_NODE_ENABLED_INSTRUMENTATIONS environment variable as documented in the Node.js zero-code instrumentation documentation.

apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
# ... other fields skipped from this example
spec:
  # ... other fields skipped from this example
  nodejs:
    env:
      - name: OTEL_NODE_ENABLED_INSTRUMENTATIONS
        value: http,nestjs-core # comma-separated list of the instrumentation package names without the `@opentelemetry/instrumentation-` prefix.

To keep all default libraries and disable only specific instrumentation libraries you can use the OTEL_NODE_DISABLED_INSTRUMENTATIONS environment variable. For details, see Excluding instrumentation libraries.

apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
# ... other fields skipped from this example
spec:
  # ... other fields skipped from this example
  nodejs:
    env:
      - name: OTEL_NODE_DISABLED_INSTRUMENTATIONS
        value: fs,grpc # comma-separated list of the instrumentation package names without the `@opentelemetry/instrumentation-` prefix.

Learn more

For more details, see Node.js auto-instrumentation.

Python

The following command will create a basic Instrumentation resource that is configured specifically for instrumenting Python services.

kubectl apply -f - <<EOF
apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
metadata:
  name: demo-instrumentation
spec:
  exporter:
    endpoint: http://demo-collector:4318
  propagators:
    - tracecontext
    - baggage
  sampler:
    type: parentbased_traceidratio
    argument: "1"
EOF

By default, the Instrumentation resource that auto-instruments Python services uses otlp with the http/protobuf protocol (gRPC is not supported at this time). This means that the configured endpoint must be able to receive OTLP over http/protobuf. Therefore, the example uses http://demo-collector:4318, which will connect to the http port of the otlpreceiver of the Collector created in the previous step.

As of operator v0.67.0, the Instrumentation resource automatically sets OTEL_EXPORTER_OTLP_TRACES_PROTOCOL and OTEL_EXPORTER_OTLP_METRICS_PROTOCOL to http/protobuf for Python services. If you use an older version of the Operator you MUST set these env variables to http/protobuf, or Python auto-instrumentation will not work.

Auto-instrumenting Python logs

By default, Python logs auto-instrumentation is disabled. If you would like to enable this feature, you must to set the OTEL_LOGS_EXPORTER and OTEL_PYTHON_LOGGING_AUTO_INSTRUMENTATION_ENABLED environment variables as follows:

apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
metadata:
  name: python-instrumentation
  namespace: application
spec:
  exporter:
    endpoint: http://demo-collector:4318
  env:
  propagators:
    - tracecontext
    - baggage
  python:
    env:
      - name: OTEL_LOGS_EXPORTER
        value: otlp_proto_http
      - name: OTEL_PYTHON_LOGGING_AUTO_INSTRUMENTATION_ENABLED
        value: 'true'

Note that OTEL_LOGS_EXPORTER must be explicitly set to otlp_proto_http, otherwise it defaults to gRPC.

Excluding auto-instrumentation

By default, the Python auto-instrumentation ships with many instrumentation libraries. This makes instrumentation easy, but can result in too much or unwanted data. If there are any packages you do not want to instrument, you can set the OTEL_PYTHON_DISABLED_INSTRUMENTATIONS environment variable.

apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
metadata:
  name: demo-instrumentation
spec:
  exporter:
    endpoint: http://demo-collector:4318
  propagators:
    - tracecontext
    - baggage
  sampler:
    type: parentbased_traceidratio
    argument: '1'
  python:
    env:
      - name: OTEL_PYTHON_DISABLED_INSTRUMENTATIONS
        value:
          <comma-separated list of package names to exclude from
          instrumentation>

Learn more

See the Python agent Configuration docs for more details.

Go

The following command creates a basic Instrumentation resource that is configured specifically for instrumenting Go services.

kubectl apply -f - <<EOF
apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
metadata:
  name: demo-instrumentation
spec:
  exporter:
    endpoint: http://demo-collector:4318
  propagators:
    - tracecontext
    - baggage
  sampler:
    type: parentbased_traceidratio
    argument: "1"
EOF

By default, the Instrumentation resource that auto-instruments Go services uses otlp with the http/protobuf protocol. This means that the configured endpoint must be able to receive OTLP over http/protobuf. Therefore, the example uses http://demo-collector:4318, which connects to the http/protobuf port of the otlpreceiver of the Collector created in the previous step.

The Go auto-instrumentation does not support disabling any instrumentation. See the Go Auto-Instrumentation repository for more details.


Now that your Instrumentation object is created, your cluster has the ability to auto-instrument services and send data to an endpoint. However, auto-instrumentation with the OpenTelemetry Operator follows an opt-in model. In order to activate automatic instrumentation, you’ll need to add an annotation to your deployment.

Add annotations to existing deployments

The final step is to opt in your services to automatic instrumentation. This is done by updating your service’s spec.template.metadata.annotations to include a language-specific annotation:

  • .NET: instrumentation.opentelemetry.io/inject-dotnet: "true"
  • Go: instrumentation.opentelemetry.io/inject-go: "true"
  • Java: instrumentation.opentelemetry.io/inject-java: "true"
  • Node.js: instrumentation.opentelemetry.io/inject-nodejs: "true"
  • Python: instrumentation.opentelemetry.io/inject-python: "true"

The possible values for the annotation can be

  • "true" - to inject Instrumentation resource with default name from the current namespace.
  • "my-instrumentation" - to inject Instrumentation CR instance with name "my-instrumentation" in the current namespace.
  • "my-other-namespace/my-instrumentation" - to inject Instrumentation CR instance with name "my-instrumentation" from another namespace "my-other-namespace".
  • "false" - do not inject

Alternatively, the annotation can be added to a namespace, which will result in all services in that namespace to opt-in to automatic instrumentation. See the Operators auto-instrumentation documentation for more details.

Opt-in a Go Service

Unlike other languages’ auto-instrumentation, Go works via an eBPF agent running via a sidecar. When opted in, the Operator will inject this sidecar into your pod. In addition to the instrumentation.opentelemetry.io/inject-go annotation mentioned above, you must also supply a value for the OTEL_GO_AUTO_TARGET_EXE environment variable. You can set this environment variable via the instrumentation.opentelemetry.io/otel-go-auto-target-exe annotation.

instrumentation.opentelemetry.io/inject-go: 'true'
instrumentation.opentelemetry.io/otel-go-auto-target-exe: '/path/to/container/executable'

This environment variable can also be set via the Instrumentation resource, with the annotation taking precedence. Since Go auto-instrumentation requires OTEL_GO_AUTO_TARGET_EXE to be set, you must supply a valid executable path via the annotation or the Instrumentation resource. Failure to set this value causes instrumentation injection to abort, leaving the original pod unchanged.

Since Go auto-instrumentation uses eBPF, it also requires elevated permissions. When you opt in, the sidecar the Operator injects will require the following permissions:

securityContext:
  capabilities:
    add:
      - SYS_PTRACE
  privileged: true
  runAsUser: 0

Troubleshooting

If you run into problems trying to auto-instrument your code, here are a few things that you can try.

Did the Instrumentation resource install?

After installing the Instrumentation resource, verify that it installed correctly by running this command, where <namespace> is the namespace in which the Instrumentation resource is deployed:

kubectl describe otelinst -n <namespace>

Sample output:

Name:         python-instrumentation
Namespace:    application
Labels:       app.kubernetes.io/managed-by=opentelemetry-operator
Annotations:  instrumentation.opentelemetry.io/default-auto-instrumentation-apache-httpd-image:
               ghcr.io/open-telemetry/opentelemetry-operator/autoinstrumentation-apache-httpd:1.0.3
             instrumentation.opentelemetry.io/default-auto-instrumentation-dotnet-image:
               ghcr.io/open-telemetry/opentelemetry-operator/autoinstrumentation-dotnet:0.7.0
             instrumentation.opentelemetry.io/default-auto-instrumentation-go-image:
               ghcr.io/open-telemetry/opentelemetry-go-instrumentation/autoinstrumentation-go:v0.2.1-alpha
             instrumentation.opentelemetry.io/default-auto-instrumentation-java-image:
               ghcr.io/open-telemetry/opentelemetry-operator/autoinstrumentation-java:1.26.0
             instrumentation.opentelemetry.io/default-auto-instrumentation-nodejs-image:
               ghcr.io/open-telemetry/opentelemetry-operator/autoinstrumentation-nodejs:0.40.0
             instrumentation.opentelemetry.io/default-auto-instrumentation-python-image:
               ghcr.io/open-telemetry/opentelemetry-operator/autoinstrumentation-python:0.39b0
API Version:  opentelemetry.io/v1alpha1
Kind:         Instrumentation
Metadata:
 Creation Timestamp:  2023-07-28T03:42:12Z
 Generation:          1
 Resource Version:    3385
 UID:                 646661d5-a8fc-4b64-80b7-8587c9865f53
Spec:
...
 Exporter:
   Endpoint:  http://demo-collector.opentelemetry.svc.cluster.local:4318
...
 Propagators:
   tracecontext
   baggage
 Python:
   Image:  ghcr.io/open-telemetry/opentelemetry-operator/autoinstrumentation-python:0.39b0
   Resource Requirements:
     Limits:
       Cpu:     500m
       Memory:  32Mi
     Requests:
       Cpu:     50m
       Memory:  32Mi
 Resource:
 Sampler:
Events:  <none>

Do the OTel Operator logs show any auto-instrumentation errors?

Check the OTel Operator logs for any errors pertaining to auto-instrumentation by running this command:

kubectl logs -l app.kubernetes.io/name=opentelemetry-operator --container manager -n opentelemetry-operator-system --follow

Were the resources deployed in the right order?

Order matters! The Instrumentation resource needs to be deployed before deploying the application, otherwise the auto-instrumentation won’t work.

Recall the auto-instrumentation annotation:

annotations:
  instrumentation.opentelemetry.io/inject-python: 'true'

The annotation above tells the OTel Operator to look for an Instrumentation object in the pod’s namespace. It also tells the Operator to inject Python auto-instrumentation into the pod.

When the pod starts up, the annotation tells the Operator to look for an Instrumentation object in the pod’s namespace, and to inject auto-instrumentation into the pod. It adds an init-container to the application’s pod, called opentelemetry-auto-instrumentation, which is then used to injects the auto-instrumentation into the app container.

If the Instrumentation resource isn’t present by the time the application is deployed, however, the init-container can’t be created. Therefore, if the application is deployed before deploying the Instrumentation resource, the auto-instrumentation will fail.

To make sure that the opentelemetry-auto-instrumentation init-container has started up correctly (or has even started up at all), run the following command:

kubectl get events -n <your_app_namespace>

Which should output something like this:

53s         Normal   Created             pod/py-otel-server-7f54bf4cbc-p8wmj    Created container opentelemetry-auto-instrumentation
53s         Normal   Started             pod/py-otel-server-7f54bf4cbc-p8wmj    Started container opentelemetry-auto-instrumentation

If the output is missing Created and/or Started entries for opentelemetry-auto-instrumentation, then it means that there is an issue with your auto-instrumentation. This can be the result of any of the following:

  • The Instrumentation resource wasn’t installed (or wasn’t installed properly).
  • The Instrumentation resource was installed after the application was deployed.
  • There’s an error in the auto-instrumentation annotation, or the annotation in the wrong spot — see #4 below.

Be sure to check the output of kubectl get events for any errors, as these might help point to the issue.

Is the auto-instrumentation annotation correct?

Sometimes auto-instrumentation can fail due to errors in the auto-instrumentation annotation.

Here are a few things to check for:

  • Is the auto-instrumentation for the right language? For example, when instrumenting a Python application, make sure that the annotation doesn’t incorrectly say instrumentation.opentelemetry.io/inject-java: "true" instead.
  • Is the auto-instrumentation annotation in the correct location? When defining a Deployment, annotations can be added in one of two locations: spec.metadata.annotations, and spec.template.metadata.annotations. The auto-instrumentation annotation needs to be added to spec.template.metadata.annotations, otherwise it won’t work.

Was the auto-instrumentation endpoint configured correctly?

The spec.exporter.endpoint attribute of the Instrumentation resource defines where to send data to. This can be an OTel Collector, or any OTLP endpoint. If this attribute is left out, it defaults to http://localhost:4317, which, most likely won’t send telemetry data anywhere.

When sending telemetry to an OTel Collector located in the same Kubernetes cluster, spec.exporter.endpoint should reference the name of the OTel Collector Service.

For example:

spec:
  exporter:
    endpoint: http://demo-collector.opentelemetry.svc.cluster.local:4317

Here, the Collector endpoint is set to http://demo-collector.opentelemetry.svc.cluster.local:4317, where demo-collector is the name of the OTel Collector Kubernetes Service. In the above example, the Collector is running in a different namespace from the application, which means that opentelemetry.svc.cluster.local must be appended to the Collector’s service name, where opentelemetry is the namespace in which the Collector resides.