How do I get rid of the GKE cluster?

[Teo12333]

hi!

In our organisation we use databricks but I do not understand why this GKE cluster keeps getting created. We deploy workspaces and compute clusters through terraform and use the GCE tag

"x-databricks-nextgen-cluster" = "true"

From my understanding, GKE is not needed for databricks to function at all, I tried manually deleting these GKE clusters and they just go recreated without any permission, yet the workspaces keep working fine and when using compute clusters within them they use GCE normally. I had this issue both when creating a new workspace through terraform and manually, it doesn't seem to be used by the previous workspace but gets deployed in the region of the latest workspace. To my understanding, this GKE cluster seems like a bug and I could not find any way to completely disable it and just use GCE for all the computing. From what I found online so far, this only seems an issue in GCP.

If this is not needed, could you please let me know how we can delete it? If its needed, could you reference the documentation article that explains how exactly its getting used? Thanks in advance!

[Louis_Frolio]

Hey @Teo12333 ,  thanks for the clear context—what you’re seeing is expected during the current GCP migration from the older GKE-based compute architecture to the newer, VM-only architecture on GCE.

What you’re seeing

• Databricks historically launched “Classic Compute” clusters on GKE in GCP, but is migrating all customer compute to GCE (VMs) and has deprecated GKE-based compute as of March 17, 2025.
• The tag x-databricks-nextgen-cluster = true routes a new cluster to the GCE-based architecture; you’ll see a “GCE” badge next to the cluster name when it’s on the new path.
• If you manually delete the system-managed GKE cluster, Databricks will re-create it until there are no GKE-based clusters still running; once idle, Databricks automatically deletes it after a default idle period (5 days). So you should not delete it manually.

Is GKE needed?

• For Classic Compute on GCP going forward, GKE is not needed—Databricks compute is moving to a pure GCE VM architecture and GKE-based compute has been deprecated.

How to stop the GKE cluster from being recreated (and remove it)

Follow these steps; once completed, the GKE cluster will auto-delete after the idle window and won’t come back:

• In the Databricks Account Console, click Update permissions to grant the new GCE permissions (or use the documented Accounts API call). This marks your workspaces “GCE-ready.”
• For any new cluster (jobs, all-purpose, SQL endpoint), add the tag x-databricks-nextgen-cluster=true during creation to force GCE now, and verify the “GCE” badge in the UI.
• Terminate any remaining GKE-based clusters (i.e., clusters without the “GCE” badge). Don’t delete the system GKE cluster itself; Databricks will auto-delete it after the idle period once there are no GKE clusters running.
• If you want to be 100% on GCE without using tags, ask to be fast-tracked—Databricks will ramp your workspace(s) so all new clusters launch on GCE by default and the GKE cluster will be removed after it’s idle.

Notes about the tag and Terraform

• The tag must be set on a new cluster at creation time; you cannot add it later to an existing interactive cluster to migrate that cluster in-place (clone or recreate it with the tag instead).
• The tag does not apply to instance pools; pools are usable by both architectures and are migrated automatically in the background. Apply the tag to the clusters that use the pools, not the pools themselves.
• Over time, once Databricks finishes ramping your workspace(s), no tag is needed—new clusters will launch on GCE by default.

References you can share

• Update permissions and “try GCE” with the x-databricks-nextgen-cluster tag; includes UI and API details.
• Migration overview for “Classic Compute on GCE” and what changes under the hood (moving off GKE to GCE VMs).
• Field guide for ramp-up and fast-track migration; explains behavior of GKE cluster deletion after idle and how to fully ramp a workspace.

Hope this helps, Louis.