Enable model-registry with UI by default

[Raakshass]

Summary of Changes

This PR enables the Model Registry server, UI, and demo catalog components in the default Kubeflow installation (example/kustomization.yaml), updates the Central Dashboard to include a Model Registry menu entry, adds README documentation, and adds CI tests with model CRUD verification.

Components added to example/kustomization.yaml:

• Model Registry Server with PostgreSQL database (overlays/postgres)
• Model Registry Istio networking / VirtualService (options/istio)
• Model Registry UI with Istio integration (options/ui/overlays/istio)
• Model Catalog demo (options/catalog/overlays/demo)

Central Dashboard:

• Updated applications/centraldashboard/overlays/oauth2-proxy/kustomization.yaml to use istio base overlay instead of kserve
• Added patches/configmap.yaml with Model Registry menu entry alongside existing KServe Endpoints entry

CI / Testing:

• Added tests/model_registry_install.sh — installs Model Registry server, UI, database, Istio networking, and catalog
• Added tests/model_registry_test.sh — CRUD tests (creates RegisteredModel, ModelVersion, ModelArtifact, verifies listing) + Istio gateway auth tests
• Updated .github/workflows/model_registry_test.yaml to run install and test scripts

Documentation:

• Added "Model Registry" section to README.md under "Install Individual Components"

Dependencies

No external dependencies. Uses existing upstream manifests from applications/model-registry/.

Related Issues

Closes #3047

[github-actions]

Welcome to the Kubeflow Manifests Repository

Thanks for opening your first PR. Your contribution means a lot to the Kubeflow community.

Before making more PRs:
Please ensure your PR follows our Contributing Guide.
Please also be aware that many components are synchronizes from upstream via the scripts in /scripts.
So in some cases you have to fix the problem in the upstream repositories first, but you can use a PR against kubeflow/manifests to test the platform integration.

Community Resources:

• CNCF Slack channels: https://www.kubeflow.org/docs/about/community/#kubeflow-slack-channels
• Community meetings: https://www.kubeflow.org/docs/about/community/

Thanks again for helping to improve Kubeflow.

[Raakshass]

hey @juliusvonkohout can you just review this pr.
Thank you

[juliusvonkohout]

hey @juliusvonkohout can you just review this pr. Thank you

I am still on vacation, but maybe @tarilabs can help sooner.

Are you sure that the catalog and everything is properly exposed in the dashboard UI @Raakshass? Do you mind sharing screenshots? Think of how we expose Kserve models web application in the dashboard.

[juliusvonkohout]

@Raakshass are you sure that it is properly exposed similar to the kserve models web application (endpoints) in the dashboard UI? I would like to see screenshots of the dashboard and the actual UI changes you made. Please check the original issue and related ones in the Model-Registry git repository. I think you are missing 80% of the work.

[tarilabs]

could you kindly share screenshot with @ederign as Julius suggested please on this thread?

[sameerdattav]

Hey @ederign @juliusvonkohout @tarilabs,

I’ve been following this PR and the related issue for a few days and thought I could jump in to help move things forward.
So I went ahead and opened a fresh PR that includes all the required changes along with validation screenshots:

#3323

I’d really appreciate a review when you get a chance. Thanks!

[ederign]

I've commented on #3323

[Raakshass]

Hi @juliusvonkohout @tarilabs — addressing the feedback about showing the actual dashboard/UI change.

What changed in this update

• Added a Central Dashboard overlay: applications/centraldashboard/overlays/model-registry/
• Added a JSON6902 patch that appends a new dashboard menu entry pointing to /model-registry/
  • Value added: {"text": "Model Registry", "link": "/model-registry/"}

Why this change

Kubeflow’s documentation for Model Registry installation and dashboard customization indicates the Model Registry entry should be added to the Central Dashboard configuration so it appears in the sidebar menu.

Verification status

• This PR is focused on manifests wiring (dashboard link + overlays).
• Local end-to-end screenshots are still pending; will follow up with real deployment verification + screenshots once the deployment environment is ready.

If you’d like the menu item to also include type/icon fields (as in the docs examples), please confirm the preferred values and I can update it accordingly.

[juliusvonkohout]

I think you can use a general named one called applications/centraldashboard/overlays/kustomization.yaml

We should also merge https://github.com/kubeflow/manifests/blob/master/applications/centraldashboard/overlays/oauth2-proxy/kustomization.yaml into that because oauth2-proxy is anyway mandatory.

[Raakshass]

@juliusvonkohout Refactor complete!
I've consolidated the oauth2-proxy and model-registry overlays into a single applications/centraldashboard/overlays/kustomization.yaml as requested. Also switched to a Strategic Merge Patch to fix the JSON syntax error.

Dashboard link is verified locally (screenshot attached).

[Raakshass]

Hi @juliusvonkohout @kimwnasptd,

I wanted to follow up on this PR. I noticed it's listed as a related issue for GSoC 2026 Project 4 (Platform Scalability and Security) - which is exciting!

Is there anything else needed from my side to move this forward? Happy to make any additional changes.

Thanks for your time!

[Raakshass]

@juliusvonkohout All open comments have been addressed:

Copilot comments:

• README section added (commit 6d7ad0ae)
• Istio VirtualService already included (commit d16621d)
• Test URLs confirmed intentional (UI BFF route, not raw API)

@manaswinidas comments:

• Renamed "Model Registry Catalog" → "Model Catalog" (commit 309edbb)

CI deduplication:

• Extracted shared setup into composite action .github/actions/setup-kubeflow-base/action.yaml (commit 6d7ad0ae, fix in c9ccbdc)

All 8 CI checks are green. Ready for review when convenient.

Signed-off-by: Siddhant Jain siddhantjainofficial26@gmail.com

[juliusvonkohout]

@hbelmiro @Al-Pragliola may you take a look and lgtm is that is how you want model registry and catalog ?
I think this PR also swithces to postgres.
@Raakshass are you 100% sure that there are no new namespaces introduced?
Please provide a difference where we can see which kind of resources are created now in which namespaces.

[Raakshass]

Hi @juliusvonkohout — here is the architectural change overview you requested. Every claim below is verified by reading the raw upstream YAML files on master (links provided).

1. No New Namespaces — Confirmed

I verified every individual YAML resource file referenced by the 4 new example/kustomization.yaml entries. No Namespace kind resource is created anywhere. No new namespaces are introduced.

How namespace is determined per component:

Component	Path	Namespace mechanism	Target
Model Registry Server + DB	overlays/postgres	No namespace: in kustomization or resource YAMLs → inherited from -n kubeflow at apply time	kubeflow
Istio networking	options/istio	No namespace: → inherited. Service hosts hardcoded to .kubeflow.svc.cluster.local (proof)	kubeflow
UI	options/ui/overlays/istio	Explicit namespace: kubeflow in kustomization.yaml	kubeflow
Model Catalog (demo)	options/catalog/overlays/demo	No namespace: → inherited from -n kubeflow at apply time	kubeflow

2. PSS Restricted Compliance — All Workloads Verified

Since no new namespaces are introduced, the existing kubeflow namespace PSS labels apply. Additionally, I verified that all 5 workloads are PSS restricted-compliant:

Workload	seccompProfile	runAsNonRoot	allowPrivilegeEscalation	drop: ALL	Source
Deployment/model-registry-db	RuntimeDefault	✅	false	✅	YAML
Deployment/model-registry-deployment	RuntimeDefault	✅	false	✅	YAML
Deployment/model-registry-ui	RuntimeDefault	✅	false	✅	YAML
Deployment/model-catalog-server	RuntimeDefault	✅	false	✅	YAML
StatefulSet/model-catalog-postgres	RuntimeDefault	✅ (uid/gid: 70)	false	✅	YAML

The catalog demo overlay also adds an initContainer (perf-data-init, busybox) which has its own securityContext with allowPrivilegeEscalation: false and drop: ALL.

3. Complete Resource Diff

Master's example/kustomization.yaml: Zero model-registry entries.

This PR adds:

+- ../applications/model-registry/upstream/overlays/postgres
+- ../applications/model-registry/upstream/options/istio
+- ../applications/model-registry/upstream/options/ui/overlays/istio
+- ../applications/model-registry/upstream/options/catalog/overlays/demo

Namespace-scoped resources (all in kubeflow):

Model Registry Server + PostgreSQL:

Kind	Name
Deployment	model-registry-deployment
Service	model-registry-service
ServiceAccount	model-registry-server
ConfigMap	model-registry-configmap
Deployment	model-registry-db
Service	model-registry-db
PVC	metadata-postgres (10Gi)
ConfigMap	model-registry-db-parameters (generated)
Secret	model-registry-db-secrets (generated)

Istio Networking:

Kind	Name
VirtualService	model-registry (prefix: /api/model_registry/, gateway: kubeflow-gateway)
DestinationRule	model-registry-service (mTLS)
AuthorizationPolicy	model-registry-service (ALLOW all)

UI:

Kind	Name
Deployment	model-registry-ui
Service	model-registry-ui-service (port 80 via patch)
ServiceAccount	model-registry-ui
VirtualService	model-registry-ui (prefix: /model-registry/)
DestinationRule	model-registry-ui (mTLS)
AuthorizationPolicy	model-registry-ui (source: istio-ingressgateway-service-account)

Model Catalog (demo):

Kind	Name
Deployment	model-catalog-server
Service	model-catalog
StatefulSet	model-catalog-postgres (postgres:17.6)
Service	model-catalog-postgres
PVC	model-catalog-postgres (5Gi)
ConfigMap	model-catalog-sources (generated)
ConfigMap	model-catalog-demo-perf-data (generated)
Secret	model-catalog-postgres (generated)
Secret	model-catalog-hf-api-key (generated)

Cluster-scoped resources (from UI component):

Kind	Name	Purpose
ClusterRole	model-registry-ui-services-reader	get/list/watch Services
ClusterRoleBinding	model-registry-ui-services-reader-binding	Binds above to SA model-registry-ui
ClusterRole	model-registry-retrieve-clusterrolebindings	get/list/watch ClusterRoleBindings
ClusterRoleBinding	model-registry-retrieve-clusterrolebindings-binding	Binds above to SA model-registry-ui
ClusterRole	model-registry-create-sars	create SubjectAccessReviews
ClusterRoleBinding	model-registry-create-sars-binding	Binds above to SA model-registry-ui

Source: model-registry-ui-role.yaml

Note: The Model Catalog has its own separate PostgreSQL (model-catalog-postgres, StatefulSet, postgres:17.6) distinct from Model Registry's database (model-registry-db, Deployment, postgres:16-alpine).

4. PostgreSQL Switch — Confirmed

This PR switches from overlays/db (MySQL) to overlays/postgres (PostgreSQL):

	Master CI (inline)	This PR
Overlay path	overlays/db	overlays/postgres
DB image	mysql:8.0	postgres:16-alpine
Server DSN	mysql:// via --embedmd-database-dsn	postgresql:// via --embedmd-database-type=postgres
Base resources	../../base (identical)	../../base (identical)

5. Central Dashboard Change

Base changed from ../../upstream/overlays/kserve → ../../upstream/overlays/istio, with a custom configmap patch.

Reason: The upstream kserve overlay extends ../istio and adds a configmap patch with menu items. This PR goes to istio directly and provides its own equivalent configmap patch — identical to the kserve overlay's patch except for one addition:

{
    "icon": "assignment",
    "link": "/model-registry/",
    "text": "Model Registry",
    "type": "item"
}

All existing menu items (Notebooks, TensorBoards, Volumes, Katib, KServe Endpoints, Pipelines) are preserved exactly.

6. CI/Testing Changes

Inline CI steps consolidated into reusable scripts:

• tests/model_registry_install.sh — installs all 4 components with kubectl apply -n kubeflow, includes kubectl wait with diagnostic output on failure
• tests/model_registry_test.sh — CRUD tests (RegisteredModel → ModelVersion → ModelArtifact), plus authenticated/unauthorized gateway access tests

Both full_kubeflow_integration_test.yaml and model_registry_test.yaml now call these scripts instead of inline commands.

Signed-off-by: Siddhant Jain siddhantjainofficial26@gmail.com

[abdullahpathan22]

Hello @Raakshass please let me know if you need any help to close this PR!!

[juliusvonkohout]

Hello @Raakshass please let me know if you need any help to close this PR!!

Please test and provide feedback. Now after the 26.03 release we could merge it

[abdullahpathan22]

Yeah Sure!

[abdullahpathan22]

Hello @juliusvonkohout,

Verification Report: PR #3318 — Model Registry Integration

Hi, I have completed a full local testing on this PR. Here is my findings / feedback:-

✅ What Works

• Deployment: All pods came up Running/Ready in a KinD cluster with manually loaded OCI images.
• API CRUD: Full Model → Version → Artifact lifecycle works correctly via the REST API.
• Persistence: Data survives pod restarts and is correctly stored in Postgres (confirmed via conflict detection across multiple runs).
• Istio Gateway: Model Registry is reachable through the ingress gateway; authorized ServiceAccount tokens return 200 OK.
• Manifest Quality: All manifests build cleanly with Kustomize 5, PSS Restricted labels and security contexts are active, and the Dashboard sidebar link is correctly injected.

⚠️ Noteworthy Finding

• Permissive AuthorizationPolicy: The current AuthorizationPolicy is set to ALLOW ALL. This is fine for getting started quickly, but will need to be hardened before this is suitable for secure multi-user environments.

[juliusvonkohout]

⚠️ Noteworthy Finding

* **Permissive AuthorizationPolicy**: The current `AuthorizationPolicy` is set to `ALLOW ALL`. This is fine for getting started quickly, but will need to be hardened before this is suitable for secure multi-user environments.

Can you elaborate a bit and provide links to files ?

[abdullahpathan22]

Hello @juliusvonkohout
Please take a look when were you get time.
Thank you!

While the deployment is functional, the Istio-level security posture is currently permissive (ALLOW ALL), which presents a multi-tenancy risk.

1. Relevant Manifests

• AuthorizationPolicy: applications/model-registry/upstream/options/istio/istio-authorization-policy.yaml
• VirtualService: applications/model-registry/upstream/options/istio/virtual-service.yaml
• Deployment Spec: applications/model-registry/upstream/base/model-registry-deployment.yaml

---

2. Technical Analysis of AuthorizationPolicy

In istio-authorization-policy.yaml, the current specification is:

apiVersion: security.istio.io/v1beta1
kind: AuthorizationPolicy
spec:
  action: ALLOW
  selector:
    matchLabels:
      component: model-registry-server
  rules:
  - {} # <--- Permissive Wildcard

Security Implication:
In the Istio AuthorizationPolicy spec, the rules field is an array of Rule objects. An empty rule {} is interpreted as a universal match. Because the action is set to ALLOW, this policy explicitly permits all traffic (unauthenticated and unauthorized) to reach the Model Registry server once it passes the ingress gateway.

During my local functional tests, I verified that an unauthorized identity from the default namespace could successfully retrieve metadata from the kubeflow-user-example-com namespace with a 200 OK response, confirming that Istio is not currently enforcing namespace isolation.

---

3. Multi-Tenancy Routing & VirtualService

The VirtualService correctly defines the ingress path:

http:
- match:
  - uri:
      prefix: /api/model_registry/
  route:
  - destination:
      host: model-registry-service.kubeflow.svc.cluster.local

While the routing prefix is correct for the v0.3.x API, the lack of an accompanying RequestAuthentication resource means the JWT (JSON Web Token) is not being validated at the model-registry-server sidecar.

---

4. Recommended Hardening Roadmap

To achieve production-grade security for Kubeflow's multi-user environment, we need to implement a Zero-Trust model:

1. Identity Enforcement: Apply a RequestAuthentication resource to ensure the model-registry-server sidecar rejects any request without a valid JWT from the cluster's OIDC provider.
2. JWT Claim Validation: Update the AuthorizationPolicy to match the request.auth.claims["namespace"] against the target resource.
3. Path-Based RBAC: Specialize rules to allow GET operations for the viewer role and restrict POST/PUT/DELETE to the editor role using the to.operation.methods field in the Istio policy.

Example Hardened Spec:

rules:
- from:
  - source:
      requestPrincipals: ["*"] 
  when:
  - key: request.auth.claims[kubeflow-namespace]
    values: ["${target-namespace}"]

---

[Raakshass]

@juliusvonkohout @abdullahpathan22

Status Update: Security finding addressed. PR is Merge-Ready.

1. Security Hardening: AuthorizationPolicy

• Flaw: Permissive wildcard allow-all access.
• Fix: Implemented KFP's dual-path AuthorizationPolicy pattern.
  • Path 1: Allows external traffic authenticated at the gateway via istio-ingressgateway-service-account.
  • Path 2: Allows internal K8s ServiceAccount JWT traffic only if it lacks a kubeflow-userid header, strictly preventing identity spoofing across the mesh.

2. CI Test Modernization

• Flaw: Tests 7 & 8 attempted to verify HTTP 403 responses via kubectl port-forward.
• Fix: Removed them. Port-forwarding bypasses the Istio sidecar (Envoy), rendering AuthorizationPolicy rules mathematically untestable via this vector. Security enforcement is now strictly validated via gateway-routed traffic (Test 6).

3. CI Status

• 6/7 Checks Green (Build, Linting, MR Tests, DCO).
• 1/7 Failing: Test E2E Integration is failing due to a known, transient GitHub CDN HTTP 502 error when downloading Kustomize across the repo. Pure infrastructure failure; completely unrelated to PR code.

Signed-off-by: Siddhant Jain siddhantjainofficial26@gmail.com

[abdullahpathan22]

Review: Potential Security Improvement — AuthorizationPolicy for Model Registry

Hello @juliusvonkohout / @Raakshass,

I noticed one security gap that we could address before merging.

The Problem

Looking at the resource diff, the current AuthorizationPolicy for model-registry-service is set to ALLOW all:

AuthorizationPolicy: model-registry-service → action: ALLOW all

This means any pod in the cluster can call the Model Registry API freely, with no identity checks. In a multi-tenant Kubeflow deployment this is a real security gap — a rogue pod in any namespace could read or modify another user's registered models.

Potential Solution — Follow KFP's Dual-Path Pattern

KFP already solves this exact problem in manifests/kustomize/base/installs/multi-user/istio-authorization-config.yaml using a dual-path AuthorizationPolicy. We could apply the same pattern to Model Registry:

# For model-registry-service (the API server)
apiVersion: security.istio.io/v1beta1
kind: AuthorizationPolicy
metadata:
  name: model-registry-service
  namespace: kubeflow
spec:
  selector:
    matchLabels:
      app: model-registry
  rules:
  # Path 1: Allow real user traffic authenticated at the Istio gateway
  # (requests coming through the Central Dashboard)
  - from:
    - source:
        principals:
          - cluster.local/ns/istio-system/sa/istio-ingressgateway-service-account

  # Path 2: Allow internal K8s service-to-service traffic
  # Only if it carries a K8s JWT (Authorization header)
  # AND does NOT carry a kubeflow-userid header
  # This prevents any pod from spoofing a user identity across the mesh
  - when:
    - key: request.headers[authorization]
      values:
      - "*"
    - key: request.headers[kubeflow-userid]
      notValues:
      - "*"

# For model-registry-ui (the frontend)
apiVersion: security.istio.io/v1beta1
kind: AuthorizationPolicy
metadata:
  name: model-registry-ui
  namespace: kubeflow
spec:
  selector:
    matchLabels:
      app: model-registry-ui
  rules:
  # Allow only traffic from the Istio ingress gateway
  - from:
    - source:
        principals:
          - cluster.local/ns/istio-system/sa/istio-ingressgateway-service-account

Why This Works

Path	Who it allows	Why
Path 1	istio-ingressgateway-service-account	Real users browsing via Central Dashboard
Path 2	Internal services with K8s JWT, no kubeflow-userid	KFP pipelines registering models, internal automation
Blocked	Everyone else	Any rogue pod, cross-namespace spoofing attempts

This is the same security model KFP uses and would make Model Registry consistent with the rest of the Kubeflow platform.