Saltar al contenido principal

MLOps Components

Updated May 14, 2023 ·

Integration with DevOps​

MLOps automates the lifecycle of machine learning models, from development to deployment and monitoring. It builds on DevOps but adds machine learning workflows and data management.

  • Uses workflows, pipelines, and artifacts.
  • Includes ML-specific components (model registry, feature store, and metadata store.)

Workflows​

Workflows are sequences of tasks that process inputs into outputs, which can be automated using pipelines.

  • Manual workflows – Require human intervention.
  • Automated workflows – Fully scripted for efficiency.
  • Pipelines – Automate workflows through structured scripts.

For more information, please see Automation Workflows.

Artifacts​

Artifacts are the outputs of pipelines which essential for deployment.

  • Catch-all term for any output of the development process
  • Include trained models, logs, and reports.

ML Pipelines​

Build pipelines convert raw code and data into deployable models. In MLOps, there are at least two separate build workflows:

  • Model build pipeline – Trains and outputs a model.
  • App build pipeline – Packages the ML application that serves the model.

Build Pipeline​

Also known as a Model Training Pipeline, the model build pipeline transforms raw data into a deployable model.

  • Features – These are raw data or pre-processed data.
  • Feature Store – A database that stores features for model training.
  • Training Output – After running, the model and its metadata are produced.

The pipeline results in a trained model and complementary data for deployment, called model metadata.

  • Model Registry – Stores and versions trained models.
  • Metadata Store – Stores model details for deployment and lifecycle management.

By organizing the model and metadata efficiently, teams can ensure smooth deployment and manage the entire lifecycle of the model.

Deployment Pipeline​

Once we have the models and app packages ready, the next step is deployment. This process involves moving the build artifacts to the target platform for serving.

  • Deployment Task – Move the ML app and model packages to the serving platform.
  • Service Start – Once deployed, start the service and monitor its performance.

After deployment, we can monitor the service to ensure it runs as expected and performs well.