Skip to Content

Langfuse Service

Langfuse  is an open-source LLM observability platform with tracing, token/cost tracking, sessions, evaluation, and prompt management. It’s available from the ARK Marketplace , and installing it wires ARK’s OpenTelemetry export to Langfuse automatically.

One backend per namespace: Langfuse and Phoenix both manage the otel-environment-variables Secret that points ARK at a backend, so install one of them per namespace. If Phoenix is already installed, uninstall it (helm uninstall phoenix -n phoenix) before installing Langfuse, or use per-tenant OTEL routing. The marketplace repo is the source of truth for install — these steps were verified against it.

Install

Langfuse is a larger stack (Postgres, ClickHouse, Redis, and S3/MinIO), so allow a few minutes for it to come up. Install with Helm from the marketplace repo:

cd services/langfuse helm repo add langfuse https://langfuse.github.io/langfuse-k8s helm dependency update chart/ helm install langfuse ./chart -n telemetry --create-namespace \ --set demo.project.publicKey=lf_pk_1234567890 \ --set demo.project.secretKey=lf_sk_1234567890

devspace deploy works too and restarts the controller for you. Installing Langfuse creates an otel-environment-variables Secret in ark-system and default pointing at Langfuse:

OTEL_EXPORTER_OTLP_ENDPOINT = http://langfuse-web.telemetry.svc.cluster.local:3000/api/public/otel

With Helm, restart the components that emit spans (DevSpace does this automatically):

kubectl rollout restart deployment/ark-controller -n ark-system kubectl rollout restart deployment/ark-completions -n ark-system

View traces

Port-forward the Langfuse UI and sign in with the demo credentials (ark@ark.com / password123), then run a query:

kubectl port-forward -n telemetry svc/langfuse-web 3000:3000 # open http://localhost:3000 ark query agent/my-agent "hello"

Under the Ark project → Tracing, each query appears as a trace:

Langfuse Tracing view listing ARK query traces

Open a trace to see the observation tree, latency, token counts, and cost (Langfuse derives these from the LLM spans), plus a flow graph:

Langfuse trace detail showing the observation tree with token counts and cost

Reference


Previous: Phoenix Service

Last updated on