Use the container registry directly

The platform embeds an OCI container registry (Docker Registry HTTP API V2) in the app. It opens at /v2 on the same address as the app, so docker · containerd · podman push and pull without a separate registry server. Training runtime images, serving images and imported partner images all live here.

Issue a token#

Permissions are checked again on every request. A VIEWER can also get a token, but that token can only pull. If the issuer's role changes later, the change applies immediately to tokens already issued. An ADMIN sees every token of the tenant in the list, other members only their own; a token is revoked by its issuer or an ADMIN.

The API issues tokens too.

curl -sS -X POST "$API/api/v1/registry/tokens" \
  -H "Authorization: Bearer $JWT" -H "X-Tenant: DEMO" \
  -H 'Content-Type: application/json' -d '{"name": "ci-push", "expires_in_days": 7}'
# -> {"id": "…", "token": "<secret>", "tenant": "DEMO", "username": "demo", "registry_host": "mlops.example.com"}

Login · push · pull#

echo "<issued token>" | docker login mlops.example.com -u demo --password-stdin

docker tag example/mnist-trainer:v1 mlops.example.com/demo/example/mnist-trainer:v1
docker push mlops.example.com/demo/example/mnist-trainer:v1
docker pull mlops.example.com/demo/example/mnist-trainer:v1

• Any user name works. The token is the password. For readability, the tenant code (lower case) is recommended.
• Do not put a scheme such as http:// in an image address. It is always host[:port]/path:tag.
• Storage is fully separated per tenant. The first path segment is just a name; whatever name you push under, the image goes into the token's tenant space. By convention, putting the tenant code in the first segment (demo/…) gives the same shape as images imported on screen.
• Another tenant's images look as if they do not exist (not found). The registry deliberately answers 404 so that "exists but no permission" cannot be told apart from "does not exist".

The standard API lists repositories and tags.

curl -u "demo:<token>" https://mlops.example.com/v2/_catalog
curl -u "demo:<token>" https://mlops.example.com/v2/demo/example/mnist-trainer/tags/list
# {"name":"demo/example/mnist-trainer","tags":["v1"]}

_catalog lists only your own tenant's repositories. Shared mlops/… images granted to you are not listed, but they pull if you know the tag.

Delete images#

/v2 has no delete request, so docker cannot delete images. Delete them in the container registry screen, or call the REST API with the 🔑 DevelopDEVELOPER · OPERATOR · ADMIN permission.

curl -sS -X DELETE -H "Authorization: Bearer $JWT" -H "X-Tenant: DEMO" \
  "$API/api/v1/registry/repositories/demo/example/mnist-trainer/tags/v1"
# {"name":"demo/example/mnist-trainer","tag":"v1","repository_deleted":true}

• Deleting the last tag removes the repository too. To delete a whole repository, use DELETE /api/v1/registry/repositories/{name}.
• If the platform uses the image (a training runtime variant, a model image, a deployment, serving and so on), the answer is 409, with the users listed in detail.references.
• Shared mlops/… images cannot be deleted.

A registry open over HTTP (insecure registry)#

If the registry is behind HTTPS, there is nothing to configure. If the platform is open only over HTTP, you must configure the client side to allow HTTP. Docker allows localhost as an exception, so it works without configuration.

{ "insecure-registries": ["mlops.example.com:10000"] }

sudo systemctl restart docker

podman login --tls-verify=false mlops.example.com:10000 -u demo
podman push --tls-verify=false mlops.example.com:10000/demo/example/mnist-trainer:v1

mirrors:
  "mlops.example.com:10000":
    endpoint: ["http://mlops.example.com:10000"]
configs:
  "mlops.example.com:10000":
    tls:
      insecure_skip_verify: true

The operator usually sets up nodes at installation. The procedure is in Installation.

The shared namespace mlops#

Repositories whose path starts with mlops/ are in a reserved shared namespace. A system administrator uses it to build training runtime · dataset stager · serving builder images once and share them with several tenants.

So sharing is per tag. Even within the same mlops/training-runtime repository, tenant A may get only the yolo-cuda tag and tenant B only the rfdetr-cpu tag. The system administrator decides who gets which shared images (Administration guide).

You cannot put images into mlops/ by import (tar upload). An import always goes into your own tenant space.

Troubleshooting#