Coder workspaces and the dashboard platform

Insights Platform

Sign in with Keycloak, get a VS Code and Jupyter workspace, publish a Streamlit, Gradio, Dash, Panel or FastAPI app to its own hostname.

Last verified 2026-11-19

  • Workspace on demand · 8h autostop
  • Profile 2–8 CPU · 2–8 GiB limits
  • Frameworks 5 + custom
  • Kube credentials none, either direction
  • Dashboard hosts one login
  • Namespaces 4

The path, as a diagram

Two lenses: the journey, and the platform it runs on.

Layout
9 components · 4 not in this view
Press enter or space to select a node. You can then use the arrow keys to move the node around. Press delete to remove it and escape to cancel.
Press enter or space to select an edge. You can then press delete to remove it or escape to cancel.
Connection types — click to isolate collected 11/20/2026, 2:44:32 AM

01 · Orientation

Two surfaces, one identity

One login for the workspace and the dashboard board.

Coder creates and runs workspaces. dash.revierie.com lists dashboards, deploys, grants and logs. Two applications and two OIDC clients, one Keycloak realm and one groups claim.

SurfaceHostnameWhat it isHow you enter
Workspacecoder.revierie.comCreates and runs the VS Code + Jupyter podOIDC client coder
Workspace apps8080--main--<workspace>--<user>.coder.revierie.comcode-server on :8080, JupyterLab on :8888The Coder session; path apps are off
Dashboard homedash.revierie.comCards, deploy, grants, logsoauth2-proxy (client dash)
A dashboard<slug>.dash.revierie.comOne app, one hostnameoauth2-proxy, then /authz/<slug>

One session covers every dashboard

oauth2-proxy sets --cookie-domain=dash.revierie.com, so one sign-in covers the home page and every subdomain. Coder keeps a separate session on the coder host.

02 · Provision

1 · Spin up a workspace

One template, three parameters, about a minute.

  1. 01

    Sign in at coder.revierie.com

    Coder redirects to Keycloak (realm homelab, client coder). There is no local account. Groups come from the token claim.

  2. 02

    Create a workspace

    Choose the Revierie DS Workspace template. CI pushes it; the image is pinned by digest.

  3. 03

    Choose the shape

    CPU 2–8 cores, memory 2–8 GiB, home disk 10–50 GiB. CPU and memory are mutable; the disk is fixed.

  4. 04

    The provisioner builds it

    A pod and Longhorn PVC appear in coder-workspaces. The pod runs as uid 1000 with automountServiceAccountToken: false and pulls the image through the Kyverno-cloned Harbor secret. Anti-affinity spreads workspaces across nodes when possible.

  5. 05

    Open an app

    The image includes VS Code (code-server 4.145.2, :8080), JupyterLab (:8888), the data-science stack and 15 extensions, including Revierie Dashboards.

  6. 06

    Stop, or wait

    Workspaces autostop after 8 hours idle and stop on disconnect. The pod is disposable; the PVC survives.

SettingValueNote
Imageds-workspace:<sha> from HarborPinned by the template; nightly rebuild with a Trivy CRITICAL gate
RuntimePython 3.12 · code-server 4.145.2 · JupyterLabnumpy, pandas, polars, matplotlib, seaborn, plotly, scikit-learn, openpyxl, duckdb
CPU / memory2–8 cores / 2–8 GiB (limits)Mutable; requests are 250m / 512Mi
Home disk10–50 GiB (Longhorn)Fixed after create; survives stops
Autostop8 hours idleAlso stopped on disconnect
Namespace quota8 CPU / 16 GiB requests, 100 Gi storageShared by all workspaces

The workspace has no cluster credentials

The pod mounts no ServiceAccount token and has no kubeconfig. kubectl and helm are on the image, but they cannot reach this cluster. The extension uses your Keycloak token, not a cluster credential.

03 · Author

2 · Build a dashboard

Any Python process that serves HTTP. Five frameworks are preinstalled.

A dashboard is a project directory: an entry file, a port, optional requirements.txt pins and a .deployignore. Run it in the workspace (streamlit run app.py, uvicorn app:api). Nothing reaches the cluster until you publish.

  • streamlit: streamlit run <entry>
  • gradio: python <entry>
  • dash: python <entry>
  • panel: panel serve <entry>
  • uvicorn: uvicorn <module>:<app>
  • custom: whatever command the deployment should run

requirements.txt is pip-installed into /app/libs on every pod boot. That is slow for heavy dependencies and repeats on restart. Bake heavy dependencies into the runtime image.

Uploads are filtered

The extension tars the project minus the patterns in .deployignore and defaults such as .git/, virtualenvs, node_modules/ and large data files. Each publish stores a new version with its sha256; the Deployment records the version it runs.

04 · Publish

3 · Publish it

The extension signs in as you. The catalog writes the objects.

  1. 01

    Sign in to the extension

    The extension runs the OAuth 2.0 device flow against the public client revierie-dash-extension. Open the link and confirm the code. Tokens go to VS Code SecretStorage and refresh on 401.

  2. 02

    Describe the dashboard

    App type, entry file, port, slug, title and access. The slug matches ^[a-z0-9][a-z0-9-]{2,19}$ and must be unique. The wizard saves ds-deploy.json.

  3. 03

    Bundle and upload

    POST /api/dashboards/<slug>/bundle sends the tar.gz with your Bearer JWT. The catalog stores the version and its sha256.

  4. 04

    The catalog generates objects

    In data-dashboards: a Deployment (Recreate, 1 replica), a Service, two forwardAuth middlewares and an IngressRoute for <slug>.dash.revierie.com.

  5. 05

    The pod fetches its code

    An init container downloads the bundle from a one-hour HMAC URL, untars it and installs requirements if present. The main container runs the pinned runtime image.

  6. 06

    Wait for the rollout

    The catalog waits three minutes for a ready pod. Recreate stops the old pod first, so a failed rollout leaves the dashboard down as deploy-failed.

  7. 07

    Open it

    The dashboard is at https://<slug>.dash.revierie.com and on the home page. The .dash cookie covers the login.

{
  "appType": "streamlit",
  "entry": "app.py",
  "port": 8501,
  "slug": "churn-monitor",
  "title": "Churn monitor",
  "description": "Weekly cohort churn, refreshed from the warehouse.",
  "access": "group:dashboard-viewers"
}
ObjectNamePurpose
Deploymentdash-<slug>1 replica · Recreate · uid 1000 · no SA token · 100m/256Mi requests, 1 CPU/1 Gi limits
Servicedash-<slug>ClusterIP on the declared port
Middlewaredash-oauth2-<slug>forwardAuth to oauth2-proxy; identity headers on
Middlewaredash-authz-<slug>forwardAuth to /authz/<slug>
IngressRoutedash-<slug>Host rule, wildcard TLS and the middleware chain

No kubectl in the extension

The extension holds only your Keycloak token. The catalog's dashboard-deployer ServiceAccount writes to data-dashboards: Deployments, Services, traefik.io Middlewares and IngressRoutes, and pod logs. It has no ClusterRoles and cannot read Secrets.

05 · Authorize

4 · Decide who can see it

Dashboards start private. Grants add users, groups or public access; every request is checked.

Grants are rows in the catalog's Postgres. An owner can share with a user, a Keycloak group, or every signed-in user. The home page filters cards by the same rule.

GranteeWhat it allowsWhere it is stored
OwnerRedeploy, stop, delete, edit details and grantsowner_email on the dashboard row
UserOpen the dashboardGrant row, grantee_type: user
Keycloak groupOpen it as any group memberGrant row, grantee_type: group, e.g. dashboard-viewers
PublicOpen it after sign-inis_public on the grant row

Traefik asks oauth2-proxy for a session. The proxy's headers carry the identity. Traefik then calls /authz/<slug>, which returns 202 for the owner or a grantee, 403 without a grant, and 404 for an unknown slug. The dashboard process has no auth code.

Anonymous traffic fails closed

The home page redirects to Keycloak. A dashboard subdomain returns 401; sign in at the home page first. Traefik cannot reference cross-namespace services in error middlewares, so the redirect lives on the base host.

Groups come from the Keycloak token, not from the cluster. Group changes apply at the next sign-in. Grant changes apply on the next request.

06 · Security

Security model

Identity, authorization, admission, isolation and supply chain are checked separately.

LayerControlEnforced by
IdentityKeycloak OIDC, realm homelab. No local passwords. Separate clients for Coder, oauth2-proxy and the extension.Keycloak
Browser authoauth2-proxy forwardAuth. HttpOnly cookie on dash.revierie.com. Identity headers only from a real login.data-platform
API authKeycloak JWT checked against the realm JWKS: signature, issuer, azp/audience, groups.catalog API
AuthorizationOwner, user, group and public grants, checked per request by /authz/<slug>.catalog API
Workload identityNo SA token in workspace or dashboard pods. The deployer SA is scoped to data-dashboards. No ClusterRoles.templates + RBAC
AdmissionKyverno ds-platform-guardrails (Enforce): uid 1000, no SA token, harbor.revierie.com/revierie/ds-* images. Workspace pods run ds-workspace only.Kyverno
IsolationBaseline Pod Security. Quota 8 CPU / 16 GiB and 100 Gi storage. LimitRange 4 CPU / 8 Gi. No NodePorts or LoadBalancers.namespaces
NetworkIngress-only NetworkPolicies: Traefik to catalog, oauth2-proxy and dashboards; Prometheus to metrics.data-platform, data-dashboards
Supply chainPrivate Harbor images, Trivy CRITICAL gate, digest and SHA pins, pull-secret cloning.CI + Harbor + Kyverno
SecretsSealedSecrets in Git. HMAC bundle tokens, one-hour TTL. The deployer Role cannot read Secrets.Git + catalog
Audit[AUDIT] lines for create, deploy, stop, start, restart and grant changes, with the acting user.catalog API → Loki

Kyverno does not know who applied an object. A generated pod and a kubectl apply face the same rules.

NetworkPolicies are ingress-only

kube-router with flannel refuses cross-node overlay traffic from any pod with an egress policy, whatever the rules allow. Probes against the API VIP and pod IPs confirmed it. Ingress is enforced; egress drafts stay in Git for a CNI that supports them.

A workspace cannot escalate through the platform. It has no token. The extension has no cluster client. Bundles arrive through a URL minted per deploy. The image allowlist rejects anything not built from this repository.

07 · Operations

Running it

Git declares the control plane. Workspaces and dashboards are runtime state.

Argo CD declares coder and data-platform from SHA-pinned Harbor images. Coder and the catalog create workspaces and dashboards at runtime; Kyverno checks them.

  • Deploy / Update: bundle the project and roll a version
  • Open: the home page or the dashboard host
  • Status: version, replica readiness, catalog status
  • Logs: the last minutes of the pod
  • Stop / Start / Restart: lifecycle without a new bundle
  • Delete: the generated objects and catalog rows
  • Manage Access: opens the grants editor
AlertMeans
DSDashboardDeployFailedRollout did not become ready. Read the card logs.
DSDashboardUnhealthyPod is crashlooping or not ready.
DSAuthzForbiddenSpikeSigned-in users are refused. Check grants and the groups claim.
DSCatalogApiMetricsDownPrometheus cannot scrape the catalog. Deploys still run.
DSWorkspaceQuotaPressurecoder-workspaces above 80% of 8 CPU / 16 GiB requests.
CoderdMetricsDownCoder metrics are not scraped.
FailureBehaviour
Bad bundle uploadRollout times out, status becomes deploy-failed, host stays down until a good bundle deploys.
Catalog API downDashboards fail closed: /authz gets no answer.
Bundle volume fullLonghorn expands the volume; an alert fires first.
Keycloak downAll SSO fails closed, including Coder.
Workspace pod lostThe PVC survives; Coder rebuilds the pod.

Longhorn keeps hourly snapshots of workspace homes and a daily R2 backup of catalog-postgres, the bundle volume and coder-postgres. Dashboard pods hold no data; the bundle is the artifact.

08 · Limits

Not in v1

Current boundaries.

  • No per-dashboard images. Dashboards run from a bundle at boot. Heavy dependencies belong in the runtime image.
  • No egress NetworkPolicies. See the CNI limitation above.
  • No GPU workspaces. A GPU profile is roadmap.
  • No per-dashboard TLS or namespace. One wildcard certificate and one data-dashboards namespace.
  • No autoscaling or blue-green. One Recreate replica: the host is down for a few seconds.
  • No automated secret rotation. The bundle HMAC and Keycloak client secrets are resealed by hand.
  • No group sync into the cluster. Group changes apply at the next sign-in.

Per-app images, GPU profiles, egress policies and per-dashboard TLS are next. The cluster map and factory cover the rest of the platform.