MLOps intro¶
MLOps is a methodology for ML engineering that unifies ML system development (the ML element) with ML system operations (the Ops element). In particular, it provides a set of standardized processes and technology capabilities for building, deploying, and operationalizing ML systems rapidly and reliably [Salama, Kazmiercczak, and Schut, 2021].
Data-centric AI¶
To get a first overview about common MLOps related issues, watch this video from AI pioneer Andrew Ng: “A Chat with Andrew on MLOps: From Model-centric to Data-centric AI”
Slides from the talk.
After watching this video, you should be able to answer the following questions:
Questions
Describe the lifecycle of an ML project.
What is the difference between a model-centric vs data-centric view?
Describe MLOps’ most important task.
Challenges¶
Despite the growing recognition of AI/ML as a crucial pillar of digital transformation, successful deployments and effective operations are a bottleneck for getting value from AI [Salama, Kazmiercczak, and Schut, 2021]:
Only one in two organizations has moved beyond pilots and proofs of concept.
72% of a cohort of organizations that began AI pilots before 2019 have not been able to deploy even a single application in production.
Algorithmia’s survey of the state of enterprise machine learning found that 55% of companies surveyed have not deployed an ML model.
Note
Most models don’t make it into production, and if they do, they break because they fail to adapt to changes in the environment.
Watch this presentation from Nayur Khan, global head of technical delivery at McKinsey, to get a first understanding of common MLOps related challenges:
After watching this video, you should be able to answer the following questions:
Questions
Describe 4 typical challenges when creating machine learning products.
Reusability concerns within a codebase: Explain a common way to look at what code is doing in a typical ML project.
What kind of problems does the open-source framework Kedro solve and where does Kedro fit in the MLOps ecosystem?
Components¶
Next, you’ll get an overview about some of the primary components of MLOps. “An introduction to MLOps on Google Cloud” by Nate Keating:
After watching this video, you should be able to answer the following questions:
Questions
Describe the challenges of current ML systems (where are teams today)?
What are the components of the ML solution lifecycle?
Explain the steps in an automated E2E pipeline.
Framework¶
To get a deeper understanding about the complete MLOps framework, read the following resources provided by Google:
MLOps: Continuous delivery and automation pipelines in machine learning
Practitioners guide to MLOps: A framework for continuous delivery and automation of machine learning from Salama et al. [2021].
After reading the reports, you should be able to answer the following questions:
Questions
Describe the difference betweeen DevOps versus MLOps
Name and explain the steps for developing ML models
Describe the three different MLOps maturity levels. In particular, explain the concepts of:
Data and model validation
Dataset and feature repository
Metadata management
ML pipeline triggers
Continuous training
Model deployment
Prediction serving
Continuous monitoring
Model governance