MLOps best practices and pitfalls

MLOps for Business

Arne Jonas Warnke

Head of Emerging Curriculum

Content of this video

MLOps

Best practices
Pitfalls to avoid

Best practices

Best practices around designing an MLOps application

Focus primarily on getting better data
- Rather than better models
- Data-centric approach

Start with business question
- Consider PoC if feasibility unclear
- Assess potential bottlenecks early on
- Ensure you learn something even when failing

A sheet with numbers

Infrastructure

Infrastructure-related best practices

Demand for computing power varies
- Especially high during training
- Infrastructure needs to be prepared for that
Keep in mind associated costs

Coworkers in server room

Best practices concerning modeling

Here are some best practices from an MLOps point of view:

Start with a simple model
- Quickly developed, can serve as baseline
- Easier to understand

When adding a new model:

Is the model better as the status quo?

A hybrid human and robot writing code

Testing

Automated testing

Across the whole life-cycle

Code
Models
Data quality
Complete pipelines

Automated Testing

Code quality and reproducibility

All results

Understandable
Reproducible

=> Testing code quality

Two programmers working together improving software code

Knowledge transfer

Versioning

Models and code
(Store) data

Maintain culture of

Documentation
Clean code
Knowledge transfer

Brainstorming image

Document insights gained

Track knowledge gained

Even if not part of the final application

A bulb

Best practices concerning operations

Automated

Logging
Monitoring practices

Team trained

To react quickly to errors

Continuous Monitoring / -Training

A monitoring room

Let's practice!

MLOps for Business