[2/2] Practitioners guide to MLOps: A framework for continuous delivery and automation of machine learning (Category ML)

Continue. first this whitepaperIt is necessary to tell about the capabilities of ML platforms that are needed in order to build good MLOps processes:

• Experimentation joint implementation of research data analysis, creation of model prototype architecture and implementation of training procedures
• Data processing The ability to process data, which allows you to prepare and convert large amounts of data in conveyors of continuous learning
• Model training The ability to efficiently and cost-effectively run powerful algorithms for training machine learning models
• Model evaluation The ability to evaluate the model in an interactive mode during experiments
• Model serving Ability to deploy and maintain models in production
• Online experimentation The ability to conduct online experiments to understand how newly trained models perform in production compared to current models (if any) before the release of a new model in production
• Model monitoring The ability to monitor models allows monitoring the effectiveness and effectiveness of deployed models to ensure predictable quality.
• ML pipelines - the ability to build complex piplanes for training and operation of models in the market
• Model registry - centralized registry of models (registration of models, description of dependencies, documentation and metadata, integration with experiments and monitoring, and roll-out and roll-back of models)
• Dataset and feature repository The storage of data sets and features allows unifying the definition and storage of data for ML models. Having a central repository of fresh, high-quality data enables sharing, discovery and reuse.
• ML metadata and artifact tracking - Filtering different types of ML artifacts that are created in different MLOps life cycle processes, including descriptive statistics and data schemas, trained models and evaluation results

As a result, combining the capabilities of platforms and understanding the target required steps of the pipeline, you can assemble an excellent process where ML will not be separated from work with data and development, which means building and operating ML systems will become easier.

P.S. Of course, reading between the lines of this document, you can see that inside the Google Cloud there is a ML platfom with all these features, which allows you to build MLOps out of the box ... if you use the right cloud:) But this does not detract from the good description of the framework for MLOps that is given in this article.

#ML #Devops #Data #AI #Software #Architecture #Processes