Maintain reliable model performance and drive continuous improvement
Standardize processes and provision tooling to optimize the deployment of models to the live environment
Streamline the development workflow to improve productivity, collaboration, and maintenance
Live model operations
Data management
Deployment
Development
¹Artificial intelligence and machine learning.
AI/ML¹ application life cycle
Enablers supporting every phase
MLOps establishes key practices across the application life cycle that increase productivity, speed, and reliability and reduce risk.
Ensuring the right talent mix and operating model to execute best practices across the AI life cycle
People
Creating reusable components to increase efficiency and reduce risk
Assetization
Establishing processes, governance, and tooling to control the AI system
Compliance, security, and risk
Provisioning the environment and tooling to optimize workflows
Technology stack
Ensure data availability, quality, and control to feed the AI system
People
Ensuring the right talent mix and operating model to execute best practices across the AI life cycle
Enabler
Assetization
Creating reusable components to increase efficiency and reduce risk
Enabler
Compliance, security, and risk
Establishing processes, governance, and tooling to control the AI system
Enabler
Technology enables learnings to be shared and collaboration across the business
Technology and tooling optimizes workflows
Teams operate inefficiently with yesterday’s tools
Fragmented technology landscape drives inconsistent standards and limits collaboration between teams
Technology stack
Provisioning the environment and tooling to optimize workflows
Enabler
Operating model
User support and feedback
Continuous improvement
Model maintenance
Monitoring
Elements
Key role
Data engineer
Data scientist
DevOps engineer
4. Live model operations
Maintain reliable model performance and drive continuous improvement
3. Deployment
Standardize processes and provision tooling to optimize the process of deploying the model to the live environment
Automated, efficient, continuous integration and delivery build testing and validation into every release
Informed production release decisions and model registry provide full transparency on production solutions
Manual and error-prone deployment with poor testing and validation
No controls for what is running in production
Model management
Model serving
End-to-end solution validation
Elements
Key role
Machine learning engineer
2. Development
Streamline the development workflow to improve productivity, collaboration, and maintenance
Solutions are built right the first time
Solutions assembled from prebuilt components and tooling enable high degree of automation
Structured and collaborative development
Highly manual work that needs refactoring before use
Individualistic, artisan, experimentation
Production-ready coding
Experiment tracking
Model choice and optimization
Elements
Key role
Data scientist
1. Data management
Ensure data availability, quality, and control to feed the AI system
Life-cycle stage
Foundations laid for reuse across hundreds of solutions with robust controls
Automated data management ensures availability of high-quality data
Massive manual effort for one-off use with no controls over quality and model drift that can damage overall performance
Data discovery
Data versioning
Data transformation
Elements
Key role
Data engineer
With MLOps
Before MLOps
With MLOps
Before MLOps
With MLOps
Before MLOps
Back
01 of 08
02 of 08
Next
With MLOps
Before MLOps
03 of 08
With MLOps
Before MLOps
The entire system is monitored with instant alerts enabling rapid issue resolution
Performance degrades over time without the business knowing about it, eroding significant value
Unstable solutions go down for weeks at a time
04 of 08
05 of 08
With MLOps
Before MLOps
Un-auditable models with hidden biases that carry unrecognized business risk in case of malfunction
Models with monitoring designed to allow audits, bias checks, and risk assessment
06 of 08
With MLOps
Before MLOps
Each developer builds components from scratch to a low standard
Where possible, teams assemble and integrate existing (internal/external) modules, rather than reinvent
07 of 08
With MLOps
Before MLOps
Academic researchers, with limited knowledge for deploying into business context
Large, highly specialized, inflexible teams try to work end-to-end without the right skill set
Software-engineering skills intrinsic to data scientist profiles
Small teams deliver much more by leveraging tools and services
People operate within a broadened, core skill set enabling specialized support
08 of 08
Life-cycle stage
Life-cycle stage
Life-cycle stage
Software engineer