Skip to content
#ML

[1/2] Practitioners guide to MLOps: A framework for continuous delivery and automation of machine learning (Category ML)

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

Regular readers of the channel may have noticed that I often talk about engineering processes and mention one way or another DevOps, SRE, Platform Engineering. But the term DevOps, which was about the convergence of development and operation, very quickly appeared similar in meaning terms, but from their fields: DevSecOps, DataOps, FinOps, ..., MLOps. More information about these metamorphoses can be found in the report.The Pipeline-Driven Organization • Roy Osherove • GOTO 2022"which I'm talking about." told. But today I wanted to talk about whitepaper 2021 A year from Google Cloud on the topic of MLOps, since we are now in the heyday of AI and ML:)

The document itself consists of three parts.

  • Overview of MLOps lifecycle and core capabilities
  • Deep dive of MLOps processes
  • Putting it all together

The first part defines MLOps.

MLOps is a methodology for ML engineering that unifies ML system development (the ML element) with ML system operations (the Ops element). It advocates formalizing and (when beneficial) automating critical steps of ML system construction. MLOps provides a set of standardized processes and technology capabilities for building, deploying, and operationalizing ML systems rapidly and reliably. And shows a clear connection between data engineering, app engineering and ml engineering. Moreover, we need built-up processes of working with data as pre-requisites for effective work on ML models, and app engineering is needed in order for trained models to work well in the market and serve their requests. The paper outlines process steps that resemble standard stories from app engineering and devops, but with little specificity.

  • ML development Experimenting with data, choosing a model and architecture, evaluating options and choosing the best one
  • Training operationalization - building a pipayne training model
  • Continuous training Continuous learning in response to new data, code changes or simply on schedule
  • Model deployment - Model deployment (It is very similar to the deployment of conventional applications.)
  • Prediction serving The model works in sales and processes requests (offline)
  • Continuous monitoring Monitoring of model operation (Here's both the usual parameters of the application, and tracking the metrics of model efficiency, data drift, changes in the distribution of the process that we predict.)
  • Data and model management It is a central, end-to-end ML artifact management function that enables auditing, tracking and compliance. This feature helps with collaboration, reuse, and the ability to detect which ML models are already in place.

Continuation about capabilties ML platforms second-part.

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