The modern MLOps framework

MLOps Deployment and Life Cycling

Nemanja Radojkovic

Senior Machine Learning Engineer

About me

Profile picture

Senior Machine Learning Engineer
Experience developing, deploying, and maintaining Machine Learning models in production
Current focus: ML deployment and life-cycling

¹ www.linkedin.com/in/radojkovic

Conceptual course

YES: High-level understanding

conceptual flowchart

NO: Hands-on skills

coding hands-on

MLOps

Machine Learning Operations

Principles
Practices
Tools

Goal: ML workflows and services

Automated
Reproducible
Integrated

ops part 1

ops part 2

mlops from dev

Chapter I

Value and necessity of MLOps

value

Model life cycle stages

life cycle

Components of the MLOps framework

building blocks

devops only

devops with mlops

High cost of ML without the Ops

USUAL STARTING POINT

Manual ML workflows
Manual deployment
Ad hoc monitoring

=> Accumulation of technical debt

Technical Debt

the implied cost of additional rework caused by choosing an easy (limited) solution now instead of using a better approach that would take longer

~ Wikipedia[1]

Famous Google paper on the topic:

"Machine Learning: The high-interest credit card of technical debt"[2]

MORE TIME AND MODELS DEPLOYED

More technical debt and model risk
Process slower, more frustrating and error prone
Ability to deliver value impaired

¹ https://en.wikipedia.org/wiki/Technical_debt ² https://research.google/pubs/pub43146/

ML workflows

Data collection and preparation
Data-labeling
Model selection
Model training
Model packaging
Model deployment
Model monitoring and maintenance

ML workflow automation == MLOps maturity

MLOps implemented

Automated
Fast
Reproducible
Explainable

we are here

non-linear

linear

spotlight 1

all eyez on me

Coming up!

Model life-cycling

life cycle

MLOps architecture

conceptual flowchart

Let's practice!

MLOps Deployment and Life Cycling