OpenTelemetry Operator for Kubernetes

The OpenTelemetry Operator is an implementation of a Kubernetes Operator.

The operator manages:

• OpenTelemetry Collector
• auto-instrumentation of the workloads using OpenTelemetry instrumentation libraries

Documentation

• API docs

Helm Charts

You can install Opentelemetry Operator via Helm Chart from the opentelemetry-helm-charts repository. More information is available in here.

Getting started

To install the operator in an existing cluster, make sure you have cert-manager installed and run:

kubectl apply -f https://github.com/open-telemetry/opentelemetry-operator/releases/latest/download/opentelemetry-operator.yaml

Once the opentelemetry-operator deployment is ready, create an OpenTelemetry Collector (otelcol) instance, like:

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

    exporters:
      logging:

    service:
      pipelines:
        traces:
          receivers: [otlp]
          processors: []
          exporters: [logging]
EOF

WARNING: Until the OpenTelemetry Collector format is stable, changes may be required in the above example to remain compatible with the latest version of the OpenTelemetry Collector image being referenced.

This will create an OpenTelemetry Collector instance named simplest, exposing a jaeger-grpc port to consume spans from your instrumented applications and exporting those spans via logging, which writes the spans to the console (stdout) of the OpenTelemetry Collector instance that receives the span.

The config node holds the YAML that should be passed down as-is to the underlying OpenTelemetry Collector instances. Refer to the OpenTelemetry Collector documentation for a reference of the possible entries.

At this point, the Operator does not validate the contents of the configuration file: if the configuration is invalid, the instance will still be created but the underlying OpenTelemetry Collector might crash.

The Operator does examine the configuration file to discover configured receivers and their ports. If it finds receivers with ports, it creates a pair of kubernetes services, one headless, exposing those ports within the cluster. The headless service contains a service.beta.openshift.io/serving-cert-secret-name annotation that will cause OpenShift to create a secret containing a certificate and key. This secret can be mounted as a volume and the certificate and key used in those receivers' TLS configurations.

Upgrades

As noted above, the OpenTelemetry Collector format is continuing to evolve. However, a best-effort attempt is made to upgrade all managed OpenTelemetryCollector resources.

In certain scenarios, it may be desirable to prevent the operator from upgrading certain OpenTelemetryCollector resources. For example, when a resource is configured with a custom .Spec.Image, end users may wish to manage configuration themselves as opposed to having the operator upgrade it. This can be configured on a resource by resource basis with the exposed property .Spec.UpgradeStrategy.

By configuring a resource's .Spec.UpgradeStrategy to none, the operator will skip the given instance during the upgrade routine.

The default and only other acceptable value for .Spec.UpgradeStrategy is automatic.

Deployment modes

The CustomResource for the OpenTelemetryCollector exposes a property named .Spec.Mode, which can be used to specify whether the collector should run as a DaemonSet, Sidecar, or Deployment (default). Look at this sample for reference.

Sidecar injection

A sidecar with the OpenTelemetry Collector can be injected into pod-based workloads by setting the pod annotation sidecar.opentelemetry.io/inject to either "true", or to the name of a concrete OpenTelemetryCollector from the same namespace, like in the following example:

kubectl apply -f - <<EOF
apiVersion: opentelemetry.io/v1alpha1
kind: OpenTelemetryCollector
metadata:
  name: sidecar-for-my-app
spec:
  mode: sidecar
  config: |
    receivers:
      jaeger:
        protocols:
          thrift_compact:
    processors:

    exporters:
      logging:

    service:
      pipelines:
        traces:
          receivers: [jaeger]
          processors: []
          exporters: [logging]
EOF

kubectl apply -f - <<EOF
apiVersion: v1
kind: Pod
metadata:
  name: myapp
  annotations:
    sidecar.opentelemetry.io/inject: "true"
spec:
  containers:
  - name: myapp
    image: jaegertracing/vertx-create-span:operator-e2e-tests
    ports:
      - containerPort: 8080
        protocol: TCP
EOF

When there are multiple OpenTelemetryCollector resources with a mode set to Sidecar in the same namespace, a concrete name should be used. When there's only one Sidecar instance in the same namespace, this instance is used when the annotation is set to "true".

The annotation value can come either from the namespace, or from the pod. The most specific annotation wins, in this order:

• the pod annotation is used when it's set to a concrete instance name or to "false"
• namespace annotation is used when the pod annotation is either absent or set to "true", and the namespace is set to a concrete instance or to "false"

When using a pod-based workload, such as Deployment or Statefulset, make sure to add the annotation to the PodTemplate part. Like:

kubectl apply -f - <<EOF
apiVersion: apps/v1
kind: Deployment
metadata:
  name: my-app
  labels:
    app: my-app
  annotations:
    sidecar.opentelemetry.io/inject: "true" # WRONG
spec:
  selector:
    matchLabels:
      app: my-app
  replicas: 1
  template:
    metadata:
      labels:
        app: my-app
      annotations:
        sidecar.opentelemetry.io/inject: "true" # CORRECT
    spec:
      containers:
      - name: myapp
        image: jaegertracing/vertx-create-span:operator-e2e-tests
        ports:
          - containerPort: 8080
            protocol: TCP
EOF

When using sidecar mode the OpenTelemetry collector container will have the environment variable OTEL_RESOURCE_ATTRIBUTESset with Kubernetes resource attributes, ready to be consumed by the resourcedetection processor.

OpenTelemetry auto-instrumentation injection

The operator can inject and configure OpenTelemetry auto-instrumentation libraries. Currently DotNet, Java, NodeJS and Python are supported.

To use auto-instrumentation, configure an Instrumentation resource with the configuration for the SDK and instrumentation.

kubectl apply -f - <<EOF
apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
metadata:
  name: my-instrumentation
spec:
  exporter:
    endpoint: http://otel-collector:4317
  propagators:
    - tracecontext
    - baggage
    - b3
  sampler:
    type: parentbased_traceidratio
    argument: "0.25"
  python:
    env:
      # Required if endpoint is set to 4317.
      # Python autoinstrumentation uses http/proto by default
      # so data must be sent to 4318 instead of 4137.
      - name: OTEL_EXPORTER_OTLP_ENDPOINT
        value: http://otel-collector:4318
EOF

The above CR can be queried by kubectl get otelinst.

Then add an annotation to a pod to enable injection. The annotation can be added to a namespace, so that all pods within that namespace wil get instrumentation, or by adding the annotation to individual PodSpec objects, available as part of Deployment, Statefulset, and other resources.

Java:

instrumentation.opentelemetry.io/inject-java: "true"

NodeJS:

instrumentation.opentelemetry.io/inject-nodejs: "true"

Python:

instrumentation.opentelemetry.io/inject-python: "true"

DotNet:

instrumentation.opentelemetry.io/inject-dotnet: "true"

OpenTelemetry SDK environment variables only:

instrumentation.opentelemetry.io/inject-sdk: "true"

The possible values for the annotation can be

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

Multi-container pods

If nothing else is specified, instrumentation is performed on the first container available in the pod spec. In some cases (for example in the case of the injection of an Istio sidecar) it becomes necessary to specify on which container(s) this injection must be performed.

For this, it is possible to fine-tune the pod(s) on which the injection will be carried out.

For this, we will use the instrumentation.opentelemetry.io/container-names annotation for which we will indicate one or more pod names (.spec.containers.name) on which the injection must be made:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: my-deployment-with-multiple-containers
spec:
  selector:
    matchLabels:
      app: my-pod-with-multiple-containers
  replicas: 1
  template:
    metadata:
      labels:
        app: my-pod-with-multiple-containers
      annotations:
        instrumentation.opentelemetry.io/inject-java: "true"
        instrumentation.opentelemetry.io/container-names: "myapp,myapp2"
    spec:
      containers:
      - name: myapp
        image: myImage1
      - name: myapp2
        image: myImage2
      - name: myapp3
        image: myImage3

In the above case, myapp and myapp2 containers will be instrumented, myapp3 will not.

Use customized or vendor instrumentation

By default, the operator uses upstream auto-instrumentation libraries. Custom auto-instrumentation can be configured by overriding the image fields in a CR.

apiVersion: opentelemetry.io/v1alpha1
kind: Instrumentation
metadata:
  name: my-instrumentation
spec:
  java:
    image: your-customized-auto-instrumentation-image:java
  nodejs:
    image: your-customized-auto-instrumentation-image:nodejs
  python:
    image: your-customized-auto-instrumentation-image:python
  dotnet:
    image: your-customized-auto-instrumentation-image:dotnet

The Dockerfiles for auto-instrumentation can be found in autoinstrumentation directory. Follow the instructions in the Dockerfiles on how to build a custom container image.

Inject OpenTelemetry SDK environment variables only

You can configure the OpenTelemetry SDK for applications which can't currently be autoinstrumented by using inject-sdk in place of (e.g.) inject-python or inject-java. This will inject environment variables like OTEL_RESOURCE_ATTRIBUTES, OTEL_TRACES_SAMPLER, and OTEL_EXPORTER_OTLP_ENDPOINT, that you can configure in the Instrumentation, but will not actually provide the SDK.

instrumentation.opentelemetry.io/inject-sdk: "true"

Target Allocator

The OpenTelemetry Operator comes with an optional component, the Target Allocator (TA). When creating an OpenTelemetryCollector Custom Resource (CR) and setting the TA as enabled, the Operator will create a new deployment and service to serve specific http_sd_config directives for each Collector pod as part of that CR. It will also change the Prometheus receiver configuration in the CR, so that it uses the http_sd_config from the TA. The following example shows how to get started with the Target Allocator:

apiVersion: opentelemetry.io/v1alpha1
kind: OpenTelemetryCollector
metadata:
  name: collector-with-ta
spec:
  mode: statefulset
  targetAllocator:
    enabled: true
  config: |
    receivers:
      prometheus:
        config:
          scrape_configs:
          - job_name: 'otel-collector'
            scrape_interval: 10s
            static_configs:
            - targets: [ '0.0.0.0:8888' ]

    exporters:
      logging:

    service:
      pipelines:
        traces:
          receivers: [prometheus]
          processors: []
          exporters: [logging]

Behind the scenes, the OpenTelemetry Operator will convert the Collector’s configuration after the reconciliation into the following:

    receivers:
      prometheus:
        config:
          global:
            scrape_interval: 1m
            scrape_timeout: 10s
            evaluation_interval: 1m
          scrape_configs:
          - job_name: otel-collector
            honor_timestamps: true
            scrape_interval: 10s
            scrape_timeout: 10s
            metrics_path: /metrics
            scheme: http
            follow_redirects: true
            http_sd_configs:
            - follow_redirects: false
              url: http://collector-with-ta-targetallocator:80/jobs/otel-collector/targets?collector_id=$POD_NAME

    exporters:
      logging:

    service:
      pipelines:
        traces:
          receivers: [prometheus]
          processors: []
          exporters: [logging]

Note how the Operator added a global section and a new http_sd_configs to the otel-collector scrape config, pointing to a Target Allocator instance it provisioned.

Compatibility matrix

OpenTelemetry Operator vs. OpenTelemetry Collector

The OpenTelemetry Operator follows the same versioning as the operand (OpenTelemetry Collector) up to the minor part of the version. For example, the OpenTelemetry Operator v0.18.1 tracks OpenTelemetry Collector 0.18.0. The patch part of the version indicates the patch level of the operator itself, not that of OpenTelemetry Collector. Whenever a new patch version is released for OpenTelemetry Collector, we'll release a new patch version of the operator.

By default, the OpenTelemetry Operator ensures consistent versioning between itself and the managed OpenTelemetryCollector resources. That is, if the OpenTelemetry Operator is based on version 0.40.0, it will create resources with an underlying OpenTelemetry Collector at version 0.40.0.

When a custom Spec.Image is used with an OpenTelemetryCollector resource, the OpenTelemetry Operator will not manage this versioning and upgrading. In this scenario, it is best practice that the OpenTelemetry Operator version should match the underlying core version. Given a OpenTelemetryCollector resource with a Spec.Image configured to a custom image based on underlying OpenTelemetry Collector at version 0.40.0, it is recommended that the OpenTelemetry Operator is kept at version 0.40.0.

OpenTelemetry Operator vs. Kubernetes vs. Cert Manager

We strive to be compatible with the widest range of Kubernetes versions as possible, but some changes to Kubernetes itself require us to break compatibility with older Kubernetes versions, be it because of code incompatibilities, or in the name of maintainability. Every released operator will support a specific range of Kubernetes versions, to be determined at the latest during the release.

We use cert-manager for some features of this operator and the third column shows the versions of the cert-manager that are known to work with this operator's versions.

The OpenTelemetry Operator might work on versions outside of the given range, but when opening new issues, please make sure to test your scenario on a supported version.

OpenTelemetry Operator	Kubernetes	Cert-Manager
v0.68.0	v1.19 to v1.25	v1
v0.67.0	v1.19 to v1.25	v1
v0.66.0	v1.19 to v1.25	v1
v0.64.1	v1.19 to v1.25	v1
v0.63.1	v1.19 to v1.25	v1
v0.62.1	v1.19 to v1.25	v1
v0.61.0	v1.19 to v1.25	v1
v0.60.0	v1.19 to v1.25	v1
v0.59.0	v1.19 to v1.24	v1
v0.58.0	v1.19 to v1.24	v1
v0.57.2	v1.19 to v1.24	v1
v0.56.0	v1.19 to v1.24	v1
v0.55.0	v1.19 to v1.24	v1
v0.54.0	v1.19 to v1.24	v1
v0.53.0	v1.19 to v1.24	v1
v0.52.0	v1.19 to v1.23	v1
v0.51.0	v1.19 to v1.23	v1alpha2
v0.50.0	v1.19 to v1.23	v1alpha2
v0.49.0	v1.19 to v1.23	v1alpha2
v0.48.0	v1.19 to v1.23	v1alpha2
v0.47.0	v1.19 to v1.23	v1alpha2
v0.46.0	v1.19 to v1.23	v1alpha2

Contributing and Developing

Please see CONTRIBUTING.md.

Approvers (@open-telemetry/operator-approvers):

• Benedikt Bongartz, Red Hat
• Yuri Oliveira Sa, Red Hat

Emeritus Approvers:

• Anthony Mirabella, AWS
• Dmitrii Anoshin, Splunk
• Jay Camp, Splunk
• James Bebbington, Google
• Owais Lone, Splunk
• Pablo Baeyens, DataDog

Maintainers (@open-telemetry/operator-maintainers):

• Juraci Paixão Kröhling, Grafana Labs
• Pavol Loffay, Red Hat
• Vineeth Pothulapati, Timescale

Emeritus Maintainers

• Alex Boten, Lightstep
• Bogdan Drutu, Splunk
• Tigran Najaryan, Splunk

Learn more about roles in the community repository.

Thanks to all the people who already contributed!

License

Apache 2.0 License.