Upgrading the chart
-
Find the chart version you run, in the
CHARTcolumn: - Read the notes below for every release between yours and the one you’re moving to, and change your values as they say.
-
Upgrade with the same values files you installed with:
Helm installs the chart’s custom resource definitions but never upgrades them. When a release changes them, apply them before you upgrade:
Upgrading to 1.0.0+
-
Breaking: the chart no longer adds the AWS Load Balancer Controller’s annotations by default. On AWS, set
ingress.aws.enabled: true. Without it, your ALBs lose their settings and health checks. -
Breaking: GKE’s settings moved from
ingress.gketoingress.gcp, and previews’ settings fromsandbartoingress.previews: -
Breaking: health checks come from
ingress.awsoringress.gcp. Each Service’salb.ingress.kubernetes.io/healthcheck-pathannotation is no longer set by default, so remove any you copied into<component>.service.annotations. -
ingress.awsandingress.gcptake the same settings:internal,redirectToHttps,timeoutSecondsandcertificates. See Networking.internal: truereplaces setting thegce-internalclass or the ALB’sscheme: internalannotation yourself, which still work. -
On AWS, the ALB’s idle timeout is now 3600 seconds (
ingress.aws.timeoutSeconds), so websockets stay open. Aload-balancer-attributesannotation of your own replaces it. -
Newer releases of the Terraform modules turn on
ingress.awsoringress.gcpin their Helm values. Setting it in your values works with any module version.
Upgrading to 0.6.0+
- Computer previews and
tembo serviceURLs are served from your install, by Sandbar, on by default. The chart adds:- An Ingress,
tembo-previews, for*.previews.<ingress.host>. - An internal load balancer,
tembo-sandbar-hosts, atsandbar.<ingress.host>, which agent hosts reach on ports 9000 and 9001.
- An Ingress,
- Previews need DNS and a certificate for
*.previews.<ingress.host>; see Networking. Without the certificate, only previews fail. On GKE, the previews’ Ingress serves thetembo-previews-tlsSecret, which cert-manager can issue. - Previews also need a Tembo release later than 0.2.892.
- To turn previews off, set
sandbar.enabled: false.
Upgrading to 0.5.0+
- Adds Sandbar, off by default, and generates the secret the API and Sandbar share (
previews.authSecret, intembo-generated). No changes needed.
Upgrading to 0.4.0+
- Adds
podAnnotations, set on every pod the chart runs. Every component runs one replica, so set your node autoscaler’s opt-out to keep it from evicting them:- GKE and the Cluster Autoscaler:
cluster-autoscaler.kubernetes.io/safe-to-evict: "false". - Karpenter and EKS Auto Mode:
karpenter.sh/do-not-disrupt: "true".
- GKE and the Cluster Autoscaler:
Upgrading to 0.3.4+
- Every component requests CPU and memory, about 1.3 vCPUs and 4.5 GiB in total, so node autoscalers size nodes for them. The API and LiteLLM get up to 5 minutes to start.
- Every pod restarts once on upgrade, including the bundled Postgres.
Upgrading to 0.3.0+
- Runs the Tembo MCP server at
/mcp, in the API pods, with atembo-mcpService and an/mcppath on the Ingress. Agents use it as theirtemboMCP server. To turn it off, setmcp.enabled: false.
Upgrading to 0.2.0+
- The chart runs the Tembo and host controller releases it was built with. Remove
image.tag,api.image.tag,web.image.tag,mcp.image.tagandcompute.controller.image.tagfrom every values file, unless you mean to pin other versions. Leaving an old override in place keeps that component on the old image even when you upgrade the chart. Remove any corresponding image digest overrides as well.