How to create a batch inference pipeline in Azure ML: step by step

João Barros 04 de November de 2026 4 min read

This tutorial shows how to create a batch inference pipeline in Azure ML to process large CSV files and generate predictions in a scalable way. A batch pipeline is useful when you have many records or want to automate periodic predictions without a real-time service.

Prerequisites

Azure account with a subscription and permissions to create resources.
Azure Machine Learning workspace created.
Azure Machine Learning SDK for Python (azure-ai-ml) installed locally or on a Compute.
Test CSV file with the same columns the model expects.
Model already registered in Azure ML (or a ready model.pkl file).

Step 1: Prepare the environment and dependencies

Why: ensuring the SDK and variables are configured prevents common errors in remote execution. Next, we authenticate and set the workspace.

pip install azure-ai-ml azure-identity pandas scikit-learn

from azure.ai.ml import MLClient from azure.identity import DefaultAzureCredential  subscription_id = "" resource_group = "" workspace_name = ""  ml_client = MLClient(DefaultAzureCredential(), subscription_id, resource_group, workspace_name)

Step 2: Package the inference script

Why: the script defines how to read the data and invoke the model. Include a minimal script that loads a registered model and applies predictions to a CSV.

# ficheiro: score.py import os import pandas as pd import joblib  def main(input_csv, output_csv):     model_path = os.environ.get("MODEL_PATH", "model.pkl")     model = joblib.load(model_path)     df = pd.read_csv(input_csv)     preds = model.predict(df)     df['prediction'] = preds     df.to_csv(output_csv, index=False)  if __name__ == '__main__':     import sys     main(sys.argv[1], sys.argv[2])

Step 3: Create a component that runs the script

Why: in Azure ML pipelines we use components to package the execution and specify requirements (CPU, image, dependencies). We create a simple YAML or define the component in Python.

from azure.ai.ml import command_component, Input, Output from azure.ai.ml.entities import Environment  comp = command_component(     name="batch-score-comp",     description="Component for batch inference from CSV",     inputs={"input_csv": Input(type="uri_file")},     outputs={"output_csv": Output(type="uri_file")},     code="./",  # directory with score.py and environment.yml     command="python score.py ${{inputs.input_csv}} ${{outputs.output_csv}}",     environment=Environment(image="mcr.microsoft.com/azureml/base:latest", conda_file="environment.yml") ) ml_client.components.create_or_update(comp)

Step 4: Register or reference the model

Why: the component needs access to the model file. You can register the model in the Workspace and mount/transfer it to the job.

from azure.ai.ml.entities import Model  model = ml_client.models.create_or_update(     Model(path="./model.pkl", name="my-model", type="mlflow" ) )

Step 5: Create and submit the batch pipeline job

Why: a job encapsulates the execution on a target compute and enables scalability. Here we configure a simple job that passes the input CSV and receives the output CSV.

from azure.ai.ml import Input, Output, command from azure.ai.ml.entities import AmlCompute  # optional: create compute compute = AmlCompute(name="cpu-cluster", size="STANDARD_DS3_V2", min_instances=0, max_instances=2) ml_client.begin_create_or_update(compute).result()  job = command(     name="batch-inference-job",     component=comp,     inputs={         "input_csv": Input(type="uri_file", path="https://.blob.core.windows.net/data/input.csv")},     outputs={"output_csv": Output(type="uri_file", path="azureml://datastores/workspaceblobstore/paths/batch-output/output.csv")},     compute="cpu-cluster" )  returned_job = ml_client.jobs.create_or_update(job) ml_client.jobs.stream(returned_job.name)

Step 6: Automate with triggers (optional)

Why: for recurring runs you can use Azure Logic Apps or Azure Data Factory to trigger the job. A common pattern is calling the Azure ML REST API to submit the job when a new file arrives in Storage.

Verify the result

Simple check: view the job logs with ml_client.jobs.stream(job.name) and inspect the output file in the datastore/Blob. The output CSV should have the 'prediction' column with generated values.

Conclusion

You created a batch inference pipeline in Azure ML, from packaging the script to submitting the job and verifying the results. Next steps: add structured logging, manage model versions and use parallelism for very large files. Tip: test locally with a subset of data before uploading to remote compute — what is the largest CSV file you have to test?