Configure Odigos with Karpenter

What is Karpenter?

Karpenter is an open-source cluster autoscaler built for Kubernetes. It dynamically launches just the right compute resources to handle your cluster’s workloads and efficiently scales them based on real-time demand. Unlike traditional autoscalers, Karpenter can provision nodes in seconds and makes scheduling decisions at the pod level, enabling highly dynamic infrastructure scaling.

Mount Method Compatibility

Odigos support for Karpenter depends on the configured mount method:

Mount method	Karpenter support	What you need
`k8s-init-container`	Works out of the box	No additional Karpenter configuration
`k8s-host-path` / `k8s-csi-driver`	Supported with Odigos Karpenter integration	Enable `karpenter.enabled` and configure NodePool `startupTaints` (see HostPath and CSI)
`k8s-virtual-device` (default)	Supported only when your Karpenter installation can apply NodeOverlay capacity	Enable the NodeOverlay feature (when required) and apply a NodeOverlay CR (see Virtual Device)

Follow the section for your mount method below.

Init Container

If you use k8s-init-container, no Karpenter-specific configuration is required. You can stop here.

Virtual Device (NodeOverlay)

With k8s-virtual-device, instrumented pods request the extended resource instrumentation.odigos.io/generic. That capacity is advertised by the Odigos device plugin on nodes where odiglet is already running. Karpenter simulates new nodes from instance-type information. Because that simulation does not include the Odigos device resource by default, Karpenter can fail provisioning with errors such as:

no instance type has enough resources ... instrumentation.odigos.io/generic: "1"

Provider and version caveats

Support depends on your Karpenter version, cloud provider, and how the controller is packaged:

NodeOverlay is an alpha Karpenter feature (available since Karpenter v1.7.x). Exact enablement steps and API readiness vary by release.
On many installs (including official AWS Karpenter Helm charts), NodeOverlay is disabled by default and must be turned on with a feature gate, for example:

helm upgrade karpenter oci://public.ecr.aws/karpenter/karpenter \
  --namespace <KARPENTER_NAMESPACE> \
  --reuse-values \
  --set settings.featureGates.nodeOverlay=true

Equivalently, set --feature-gates NodeOverlay=true / FEATURE_GATES=...,NodeOverlay=true on the Karpenter controller.

Not every Karpenter distribution applies NodeOverlays during provisioning, even if the CRD and feature gate exist. Validate in your environment that creating the overlay below allows Karpenter to launch nodes for pods that request instrumentation.odigos.io/generic.

If NodeOverlay is unavailable or ineffective in your Karpenter install, use k8s-init-container, or k8s-host-path / k8s-csi-driver with the startup taint integration.

Apply a NodeOverlay

After NodeOverlay support is enabled in your Karpenter controller, apply a NodeOverlay that adds the Odigos device capacity for the NodePool that should run instrumented workloads. Replace default with your NodePool name if needed:

cat <<'EOF' | kubectl apply -f -
apiVersion: karpenter.sh/v1alpha1
kind: NodeOverlay
metadata:
  name: odigos-instrumentation
spec:
  requirements:
    - key: karpenter.sh/nodepool
      operator: In
      values: ["default"]
  capacity:
    instrumentation.odigos.io/generic: "1024"
EOF

Confirm the overlay is ready:

kubectl get nodeoverlay odigos-instrumentation

Karpenter should then be able to provision nodes for pods that request instrumentation.odigos.io/generic. After the node registers, odiglet advertises the real device capacity so pods can schedule.

HostPath and CSI (Startup Taints)

For the k8s-host-path and k8s-csi-driver mount methods, Odigos adds a node affinity rule to instrumented pods so they only run on nodes where odiglet is installed. This works fine in static clusters — but in Karpenter-managed clusters, this causes a problem:

odiglet is installed by a DaemonSet after the node is created.
The label odigos.io/odiglet-installed=true is only added to the node at runtime.
Karpenter does not know in advance that this label will appear.
So when the scheduler tries to place a pod with this affinity, Karpenter concludes: “no nodes exist (or can be created) that match this label” — and no node is provisioned.

Once you enable the Karpenter integration option (explained below), Odigos will no longer apply node affinity to new instrumented pods. Instead, it relies on a startup taint mechanism:

Nodes launched by Karpenter should be initialized with a special taint (odigos.io/needs-init=NoSchedule).
The exact method for configuring this taint in Karpenter will be explained in the following section.
The odiglet removes this taint after it prepares the node for instrumentation.
This allows instrumented pods to schedule only after the node is ready — achieving the same safety as affinity, without blocking provisioning.

1. Configure the Karpenter NodePool CRD

You must configure your Karpenter NodePool to include a startup taint. This ensures Odigos can prepare each node before instrumented pods are scheduled onto it. Add the following under spec.template.spec.startupTaints:

spec:
  template:
    spec:
      startupTaints:
      - key: odigos.io/needs-init
        effect: NoSchedule

To safely append this taint without overwriting any existing startupTaints, use the following command:

export NODEPOOL_NAME=<NODEPOOL_NAME>
kubectl patch nodepool $NODEPOOL_NAME --type merge --patch "$(kubectl get nodepool $NODEPOOL_NAME -o json | jq -c '.spec.template.spec.startupTaints = (.spec.template.spec.startupTaints // []) + [{"key":"odigos.io/needs-init","effect":"NoSchedule"}] | {spec: {template: {spec: .spec.template.spec}}}')"

Replace <NODEPOOL_NAME> with the actual name of your Karpenter NodePool.

2. Add the Startup Taint to Existing Nodes

To manually add the startup taint to all current nodes:

kubectl get nodes -o jsonpath='{range .items[*]}{.metadata.name}{"\n"}{end}' | xargs -I {} kubectl taint node {} odigos.io/needs-init=:NoSchedule --overwrite

This ensures existing nodes behave consistently until they’re replaced by Karpenter-managed ones.

3. Enable the Karpenter Integration in Odigos

Once the tainting setup is complete, inform Odigos to disable node affinity and rely on the taint mechanism.

Odigos CLI
Helm Chart

Run the following using the Odigos CLI:

odigos set config karpenter-enabled=true

If installing or upgrading using helm:

helm upgrade --install odigos odigos/odigos \
  --namespace odigos-system \
  --set karpenter.enabled=true

4. Restart the Odiglet DaemonSet

To ensure that taint-removal logic is triggered on all nodes, perform a rolling restart of the odiglet DaemonSet:

kubectl rollout restart daemonset/odiglet -n odigos-system

Getting Started

Odigos AI

Instrumentations

Odigos Pipeline

Destinations

Concepts

Authentication

Contribution Guidelines

Configure Odigos with Karpenter

What is Karpenter?

Mount Method Compatibility

Init Container

Virtual Device (NodeOverlay)

Provider and version caveats

Apply a NodeOverlay

HostPath and CSI (Startup Taints)

1. Configure the Karpenter NodePool CRD

2. Add the Startup Taint to Existing Nodes

3. Enable the Karpenter Integration in Odigos

4. Restart the Odiglet DaemonSet

​What is Karpenter?

​Mount Method Compatibility

​Init Container

​Virtual Device (NodeOverlay)

​Provider and version caveats

​Apply a NodeOverlay

​HostPath and CSI (Startup Taints)

​1. Configure the Karpenter NodePool CRD

​2. Add the Startup Taint to Existing Nodes

​3. Enable the Karpenter Integration in Odigos

​4. Restart the Odiglet DaemonSet

What is Karpenter?

Mount Method Compatibility

Init Container

Virtual Device (NodeOverlay)

Provider and version caveats

Apply a NodeOverlay

HostPath and CSI (Startup Taints)

1. Configure the Karpenter NodePool CRD

2. Add the Startup Taint to Existing Nodes

3. Enable the Karpenter Integration in Odigos

4. Restart the Odiglet DaemonSet