MaaS Administrator Guide

This guide explains how platform administrators manage Model as a Service (MaaS): import model services, create published models with one or more upstreams, create subscriptions, enforce token quotas, configure optional guardrails, and review usage.

For the product overview, see Model as a Service (MaaS). For the developer workflow, see MaaS User Guide.

Make a platform inference service available to MaaS

In Alauda AI Platform 2.9, an inference service is an upstream candidate; it is not itself a published MaaS model. InferenceService and LLMInferenceService are supported.

  1. Open the project or namespace that contains the inference service.
  2. Open the service management page.
  3. Select Enable for MaaS from the service actions to set maas.alauda.io/available: "true".

The service then becomes selectable as an upstream from MaaS → Models → Published Models. To remove it from future selection, set the availability label to "false". A referenced service cannot be disabled until it is removed from every published model that uses it.

This action does not create or delete a MaaSModelRef. Published models, their client names, and their routing configuration are managed centrally from the Published Models page.

Import an external model service

Use an external model service when the model is hosted outside the platform or is not represented by an InferenceService or LLMInferenceService. The 2.9 console imports OpenAI-compatible endpoints with an optional Bearer API key; this is not a native Azure, Bedrock, or GCP credential configuration workflow.

  1. In the Administrator view, open MaaS → Models.
  2. On the Model Services tab, click Import External Model Service.
  3. Enter a lowercase name for the service.
  4. Enter the provider's OpenAI-compatible endpoint, including its API base path when required. For example, use https://host/v1 when the provider expects /v1 requests.
  5. Enter an upstream API key if the provider requires one. Leave it empty for an endpoint that does not require credentials.
  6. Click Create.

The API key is stored in a Kubernetes Secret and injected by the gateway when it calls the provider. The consumer's MaaS API key and the provider's upstream key are separate credentials: developers receive only the MaaS key.

Create a published model

After making a self-hosted service available to MaaS or importing an external service, create the published model centrally. The old service-row Publish Models workflow has been removed.

  1. Open the Published Models tab.
  2. Click Add Model.
  3. Enter the client-facing model name.
  4. Select a MaaS-enabled InferenceService or LLMInferenceService, or an imported AIServiceBackend. For an external service, select its provider model name; for self-hosted services, the serving model name is discovered from the service.
  5. Use Add Upstream to add other compatible self-hosted or external services when needed.
  6. Configure priority, weight, and retry/failover settings when required.
  7. Save the published model.

One published model can combine multiple providers or self-hosted services. For example, the client can call team-chat while a self-hosted Qwen service is the primary upstream and an OpenAI-compatible provider serves provider-chat-model as a fallback. Choose upstreams with compatible request and response capabilities. The Published Models list shows the client model name, upstream count and readiness, and overall phase.

Configure upstream priority, weight, and retry

  • Use priority: 0 for the preferred upstream and larger values for fallback tiers.
  • Use weight to distribute traffic among upstreams with the same priority. It is a relative value, not an independent percentage.
  • Configure retry triggers and retry limits on the published model when a request should fail over within the same request.
  • Retry does not permanently change an upstream's priority or health state. A streaming response cannot move to another upstream after the response has started.
  • Retrying can cause a provider to process a request more than once; use it only when the request semantics allow it.

Create a subscription

A subscription is the unit of model access and token quota. A user must be a subscriber of a subscription that contains a model before they can call that model.

  1. In the Administrator view, open MaaS → Subscriptions.
  2. Click Create Subscription.
  3. Enter a subscription name.
  4. Add one or more subscribers by username.
  5. Add one or more published models. A model is selected by its client-facing model name; the same client model name cannot be added twice to one subscription.
  6. For each model, configure at least one token limit and time window.
  7. Optionally add more quota tiers for the same model.
  8. Click Save.
INFO

Users can see a subscription in My Subscriptions only when they are included as subscribers. Add each user's username to Subscribers to grant access to that user. To make the subscription available to all authenticated users, add * as a subscriber. The * value is a wildcard, not a literal username; users still need a valid MaaS API key to call the models.

For each model, the form accepts a limit in millions of tokens and a window such as Minute, Hour, or Day. For example, you can set 100M tokens per hour and 1000M tokens per day. All configured tiers are enforced.

A subscription can contain several models, and each model can have different limits. This lets you provide different access tiers without creating a separate gateway for each model. If one published model has multiple upstreams, priority and weight are configured on the published model, not on the subscription.

The subscription list shows the subscription name, models, subscribers, and current phase. Use the row action menu to view, edit, or delete a subscription.

Configure guardrails

MaaS can optionally check model requests and responses with configured content guardrails.

Enable guardrails

  1. Open MaaS → Guardrails.
  2. Click Edit settings.
  3. Enable guardrails.
  4. Select a failure policy:
    • End request on guardrails failure (FailClosed) rejects or ends the request when the guardrails engine cannot complete the check.
    • Ignore guardrails on failure (FailOpen) allows the request to continue when the guardrails engine fails.
  5. Save the settings.

Disabling guardrails removes the data-plane integration. Guardrail settings that remain on subscriptions do not take effect while the global feature is disabled.

Create a guardrail configuration

  1. On the Guardrails page, click Create Config.
  2. Enter a configuration ID using lowercase letters, numbers, and hyphens.
  3. Configure the guardrail models, including model type, model name, and base URL.
  4. Add an optional API key for a guardrail model. Credentials are stored as Kubernetes Secrets and exposed to the guardrails engine through environment variables.
  5. Edit the config.yaml content.
  6. Edit the rails.co content when your configuration requires Colang flows.
  7. Click Save.

A configuration can contain multiple guardrail models, but each model type must be unique within the configuration. The Guardrails list shows whether a configuration is attached to a subscription.

Attach guardrails to a subscription model

When creating or editing a subscription, enable guardrails for a model and select one of the available guardrail configurations. The setting is applied per model, so different models in the same subscription can use different configurations or no guardrails.

Review usage and reports

Subscription management usage

The Subscriptions page shows the subscriptions and their configured limits. Use the subscription's usage view, where available, to review token consumption for the selected subscription and model.

Usage report

Open MaaS → Report to review administrator-level usage:

  1. Select a time range: Last 1 day, Last 7 days, Last 30 days, Last 90 days, or Last 1 year.
  2. Optionally select a subscription.
  3. Search by username.
  4. Click refresh to reload the data.

The report displays:

  • Username
  • Subscription
  • Client model
  • Upstream model
  • Total tokens
  • Input tokens
  • Cached tokens
  • Output tokens
  • Reasoning tokens

Click Export CSV to export the filtered report. The export dialog lets you group data by subscription, user, API key, client model, upstream model, or legacy model, and group time by day or month.

Usage data is intended for operational visibility, quota observation, and showback. It is not a billing settlement record.