# Azure ML Batch Endpoint

> Azure ML batch endpoints are a free, stable interface for batch scoring, and cost accrues only when a scoring job pulls nodes from a compute cluster. ZopNight discovers each endpoint with its deployment linkage, mapping scoring workloads to the clusters they consume so batch spend lands on the right owner.

Source: https://zop.dev/integrations/azure/ml-batch-endpoint
Updated: 2026-08-19

---

Batch endpoints provide a stable interface for running batch scoring jobs against registered models. The endpoint itself is free; its jobs consume compute-cluster nodes on demand.

## Zero standing cost, on-demand consumption

Unlike its real-time sibling, a batch endpoint holds no provisioned instances at all. Nothing bills until a scoring job is invoked, at which point the job draws nodes from the compute cluster named by the active batch deployment, pays for them for the duration of the run, and releases them. The endpoint's contribution to the bill is exactly the runs invoked through it. An idle batch endpoint is genuinely free in a way almost nothing else on this surface is.

## Linking scoring traffic to cluster spend

Discovered via the AML enricher with deployment linkage, mapping batch-scoring workloads to the compute clusters they consume. This is attribution work: cluster node-hours are the visible charge, and the endpoint linkage explains which scoring pipeline caused them, so a spike on a shared cluster can be traced to the batch workload behind it.

## Nothing to schedule, something to audit

With no standing capacity, there is no stop verb, so ZopNight scopes batch endpoints to discovery and topology. The cost questions live downstream: whether the target cluster's minimum node count fits the scoring cadence, and whether the scoring jobs themselves run at a sensible size and frequency.

## How batch scoring still overspends

The leaks are behavioral. Scoring pipelines scheduled hourly against data that changes daily multiply node-hours 24-fold for identical output. Runs pointed at oversized clusters finish barely faster while billing bigger nodes. And endpoints abandoned by their pipelines linger harmlessly: free, but noise that hides the endpoints still doing work.

## Batch endpoints in ML studio

Azure ML studio → Endpoints → Batch endpoints lists each endpoint, its default deployment, and its job history; the job list is where the actual money shows up.