Skip to content

Latest commit

 

History

History

helm

Helm

Congratulations! You have successfully found the Helm examples. These examples are meant for advanced users looking to deploy Tempo in a microservices pattern. If you are just getting started might I recommend the docker-compose examples. The docker-compose examples also are much better at demonstrating trace discovery flows using Loki and other tools.

If you're convinced this is the place for you then keep reading!

Initial Steps

To test the Helm example locally requires:

  • k3d > v3.2.0
  • helm > v3.0.0

Create a cluster

k3d cluster create tempo --api-port 6443 --port "3000:80@loadbalancer" --agents 2

Next either deploy the microservices or the single binary.

Microservices

The microservices deploy of Tempo is fault tolerant, high volume, independently scalable.

Note: double check you're applying to your local k3d before running this!

helm repo add grafana https://grafana.github.io/helm-charts
helm repo update

Install Tempo, Grafana and run a K6 job against tempo.

helm upgrade -f microservices-tempo-values.yaml --install tempo grafana/tempo-distributed
helm upgrade -f microservices-grafana-values.yaml --install grafana grafana/grafana
kubectl apply -f microservices-extras.yaml

Single Binary

** Note: This method of deploying Tempo is referred to by documentation as "monolithic mode" **

The Tempo single binary configuration is currently setup to store traces locally on disk, but can easily be configured to store them in an S3 or GCS bucket. See configuration docs or some of the other examples for help.

Note: double check you're applying to your local k3d before running this!

helm repo add grafana https://grafana.github.io/helm-charts
helm repo update

Install Tempo, Grafana and synthetic-load-generator

helm upgrade --install tempo grafana/tempo
helm upgrade -f single-binary-grafana-values.yaml --install grafana grafana/grafana
kubectl create -f single-binary-extras.yaml

Find Traces

Navigate to http://localhost:3000/explore and try a simple TraceQL query like {}

Clean up

k3d cluster delete tempo