values-byoc.yaml, installs a data plane in a cluster and
cloud account you own and connects it to your Dippr tenant. It turns on the full plane (data apps,
the app registry, agents, pipelines, workspace databases, the sample dataset and engine metrics) and
asks you for every value that is yours: your domain, your bucket, your identities and your secrets.
A customer-operated data plane is a data plane whose cluster, storage, DNS and certificates belong
to you. Your query results, app content, workspace database rows and agent conversations stay in
your account; metadata and query text are sent to Dippr, as listed in
What leaves your data plane.
Dippr schedules work onto the plane and the plane calls out to Dippr, never the reverse.
Register the plane first (Register a data plane): the registration
returns the plane id, its queue name and the plane credential you put in a Secret below.
Install it
Layer three files: your cloud’s file, the profile, and one file of your own with the values listed in the next section.values-aws.yaml instead of values-gcp.yaml on AWS. Azure is not supported by this profile yet:
the plane cannot reach Azure storage for files, workspace databases and the sample dataset, so the
install stops with a message saying so. Do not add
values-remote.yaml: the profile already carries its settings.
If a value is missing, the install stops before anything is created and the message names the value:
Values you supply
Every value below is required. Put them inmy-plane.yaml.
On AWS, also name the plane’s bucket role under
eks.amazonaws.com/role-arn in both
worker.serviceAccount.annotations and wsdb.serviceAccount.annotations (workspace databases run as
their own service account). The example
value in the cloud file does not count: the install stops until you name your own. On Google Cloud, set objectStore.gcpServiceAccount to the Google
service account the plane acts as (see Object storage below). The plane needs one there because file
uploads and workspace database loads use signed URLs, which Google signs only for a service account.
Secrets you create
The profile renders no Secret from your values file except the two registry tokens. Create these in the plane’s namespace and name each one in its value:
A value such as
secrets.objectproxy.secret set inline is refused, so a credential never sits in a
values file.
Optional values
Prerequisites
A cluster that enforces network policies
Data apps and agents run code you install, and network policies are what keep each one to its own connections. Use a network plugin that enforces them (Calico, Cilium, Antrea, or your cloud’s policy add-on), and confirm it before you setappgw.requireEnforcingNetworkPolicy: true.
cert-manager
Engines use a certificate in a combined format, which cert-manager issues through itsAdditionalCertificateOutputFormats feature. cert-manager v1.16.2 issues it with no extra settings.
With an older release, turn the feature gate on for both the controller and the webhook:
Ready.
The apps namespace, created by a cluster admin
The plane runs data apps in a namespace calledapps. The profile does not create it, so the person
installing the plane needs no cluster-wide permission. A cluster admin creates it once, with these
labels:
restricted labels make your cluster refuse any privileged pod in that namespace. Keep them.
What the installer needs
The installer needs the built-inadmin role in two namespaces, and nothing cluster-wide:
roles/editor project role is not enough: it cannot create role bindings, so
installing cert-manager or an ingress controller fails. Those need a cluster admin.
Every pod the profile installs meets the restricted pod security level, with one exception: the
registry’s garbage collection runs as a Job the plane starts later, and that Job does not meet
restricted, because it must be able to delete any file the registry wrote. If the plane’s namespace
enforces restricted, your cluster refuses that Job, the registry never reclaims the space of deleted
data app images, and its volume keeps growing. So enforce baseline on the plane’s namespace and keep
restricted as a warning and in the audit log:
Nodes pull data app images
Data apps run images stored in the plane’s own registry, and your nodes pull them from the address you set inappRegistry.store.pull.clusterIP. Configure each node’s container runtime (containerd)
to use it. With config_path = "/etc/containerd/certs.d" in containerd’s configuration, write this
file on every node, replacing altimate with the plane’s namespace and 10.96.42.5 with your address:
appRegistry.store.pull.containerdBootstrap: true, which writes the file
from a privileged pod on each node in the plane’s namespace. That pod uses the host’s process
namespace and a host directory, which the baseline level refuses too, so it runs only if the plane’s
namespace enforces privileged. With the labels above, write the file yourself instead.
Object storage
Create a bucket for this plane alone and grant it to the plane:
Without
roles/storage.legacyBucketReader, or with soft delete on, the plane can never confirm a
deleted tenant’s files are gone.
Agents need a second, narrower identity, the one you name in appgw.serviceAccount.annotations. It
may only list, read and create objects under _altimate/agents/ in the same bucket (and replace the
inbox objects there), and must not be the plane identity above:
On both clouds, give the agents identity the same access under
_altimate/routine-guidance/, where
agents keep the guidance owners give their routines: on Google Cloud, a second managed folder with
the same two roles; on AWS, the same statements for that prefix. Without it, saving guidance fails.
Outbound access
The plane calls your region over HTTPS (443) and the work endpoint (7233). Give the plane a stable outbound address, for example private nodes behind a NAT with a reserved IP, and declare it when you register the plane: registration refuses a BYOC plane with no outbound IPs. Once your region restricts plane connections, it admits the plane only from those addresses.Uninstall the plane
helm uninstall leaves four objects in place on purpose. Deleting a tenant has to remove that
tenant’s data apps, and these objects are how the plane remembers where apps ran:
If you change
appgw.appsNamespace, first upgrade to this version with the old name and let the
plane’s worker start. Then rename in a second upgrade. The worker has then recorded the old namespace,
so deleting a tenant still removes apps left there. helm upgrade refuses a rename that would lose
it. If you render with helm template or a GitOps tool such as Argo CD, nothing checks this for you.
Remove them only after you have retired the plane and deleted its tenants:
apps, ask them to delete it, or to delete just the worker’s access with
kubectl -n apps delete rolebinding,role worker-app-lifecycle. Until the RoleBinding is deleted, a
ServiceAccount named worker that you create again in the plane’s namespace, for example by
reinstalling the plane there, gets the same access to apps.