Skip to content
CSA Loom — the Microsoft Fabric experience for Azure tenants where Fabric isn't yet available: lakehouses, warehouses, notebooks, semantic models, Activator rules, Data Agents, across Commercial, GCC, GCC-High, and DoD IL5

Tutorial: Model serving endpoint editor

CSA Loom model-serving-endpoint editor — deploy a registered model behind an HTTPS scoring endpoint on Azure ML managed online endpoints (the Azure-native default) or Databricks Mosaic AI Model Serving (opt-in via LOOM_MODEL_SERVING_BACKEND=databricks). One-for-one with the Azure ML Endpoints / Databricks Serving experience. No Microsoft Fabric.

What it is

The last mile of an ML workflow: take a registered model version, put it behind a managed HTTPS endpoint, canary a new version with a blue/green traffic split, score it from a console, and watch real Azure Monitor metrics — all without leaving Loom.

Four tabs:

Tab What it does
Overview Backend badge, live endpoint list, select / bind, endpoint state, per-row Traffic and Delete
Deployments & traffic Create an endpoint, the deployments table, and the traffic-split dialog
Invoke Scoring console — a real POST with measured round-trip latency
Monitoring Live latency / requests / error tiles and charts from Azure Monitor

When to use it

  • A registered model needs a production HTTPS scoring surface.
  • You want to canary a new model version at 10% traffic before promoting it.
  • You need latency and error-rate telemetry attached to the endpoint you own.

Step-by-step in Loom

  1. Create the item. + New item → Model serving endpoint. The header badge names the active backend (Azure ML managed online endpoints by default).
  2. See what already exists. Overview lists live endpoints read from the real backend, with their state. Click a row to select it — that selection drives the Invoke and Monitoring tabs. Each row also offers Traffic and Delete.
  3. Create an endpoint. On Deployments & traffic → Create a serving endpoint:
  4. Endpoint name (for example fraud-scorer).
  5. Registered model — a dropdown of your real registered models (with the latest version pre-selected); falls back to a text input if the registry could not be listed.
  6. Version.
  7. Instance type (Azure ML, for example Standard_DS3_v2) or Workload size (Databricks, for example Small).
  8. Scaling — Manual (fixed Instances) or Autoscale (Min / Max instances).
  9. Scale to zero when idle — a switch, Databricks backend only.

Create endpoint deploys the model version and routes 100% of traffic to the first ("blue") deployment. 4. Review the deployments. The deployments table lists deployment name, model, compute (type ×count), scale type, current traffic percentage, and state. 5. Canary a new version. Add a second deployment, then Split traffic. The dialog gives one slider per deployment in 5% steps with a running total badge; Apply split is enabled only when the total is exactly 100% and applies a real backend update (blue/green canary). 6. Score it. On Invoke, edit the JSON body to match your model's signature and click Invoke. Loom sends a real scoring request and shows the HTTP status badge, the measured round-trip latency in ms, and the raw response body. 7. Watch it run. Monitoring shows three tiles — average request latency, requests per minute, and 5xx errors per minute (red when non-zero) — plus time-series charts, sourced live from Azure Monitor (Microsoft.MachineLearningServices/workspaces/onlineEndpoints: RequestLatency, RequestsPerMinute). Refresh re-reads them.

The Azure backend it rides on

  • Serving (default): Azure ML managed online endpoints — real ARM create/update, deployments, and traffic routing.
  • Serving (opt-in): Databricks Mosaic AI Model Serving, selected with LOOM_MODEL_SERVING_BACKEND=databricks.
  • Metrics: Azure Monitor metrics for the online endpoint resource.
  • Routes: GET/POST/DELETE /api/items/model-serving-endpoint/<id>, …/traffic, …/invoke, …/metrics — every control calls a real BFF route.

Honest gates

Condition What you see Exact remediation
No serving backend configured The shared Fix-it gate (svc-model-serving) with an inline wizard; the whole surface still renders and the create form's inputs are disabled — no red banner on a fresh item Set the variable the gate names via its Fix it wizard
Endpoint has no deployments yet "No deployments yet — the endpoint may still be provisioning"; Split traffic is disabled Wait for provisioning, or create a deployment
Traffic percentages do not total 100 Apply split stays disabled and the total badge turns amber Adjust the sliders until the total is 100%
Endpoint-level metrics unavailable (Databricks path) An info MessageBar stating the reason honestly, instead of empty charts Use the backend's own serving telemetry, or switch to the Azure ML path
No endpoint selected Invoke / Monitoring prompt you to select one Pick an endpoint on Overview

No Fabric required

Azure ML (or Databricks) + Azure Monitor. No Fabric capacity, workspace, OneLake path, or Power BI workspace is used on any path.

Learn more