Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
SlideShare a Scribd company logo

What is MLOps

•
•

A brief introduction to MLOps (machine learning operations). For more, check out our other MLOps materials on valohai.com/mlops

1 of 16
Download to read offline
What is MLOps? V 0.1
What ML development should look like
Design 10% Model Development 90% Glue Coding Complex & unstructured handover Operations Time to Market What it actually looks like
Why? ML Code Needed to prove value
Why? Because most of your ML system is not about the ML code. Data Verification Serving Infrastructure Configuration Data Collection Feature Extraction Process Management Tools Analysis Tools Monitoring Machine Resource Management ML Code Hidden Technical Debt in Machine Learning Systems (2015): https://papers.nips.cc/paper/2015/file/86df7dcfd896fcaf2674f757a2463eba-Paper.pdf Needed to materialize value
The goal of MLOps is to reduce technical friction to get the model from an idea into production in the shortest possible time with as little risk as possible.
MLOps is not a single tool or platform. Tooling
MLOps is about agreeing to do ML the right way and then supporting it. Process Tooling
Process A few shared principles will take you a long way… ML should be collaborative. ML should be reproducible. ML should be continuous. ML should be tested & monitored.
Tooling And tooling will help implement your process. ML should be collaborative. ML should be reproducible. ML should be continuous. ML should be tested & monitored. Shared Infrastructure Versioning for Code, Data and Metadata Machine Learning Pipelines Model Deployment and Monitoring
Time Risk Short-term thinking Long-term thinking Do it fast Do it safe Time Do it right So it’ll be faster next time Risk Avoid the “It’s too early for MLOps” trap.
Data analysis ML pipeline Model serving Model monitoring Experimentation Feature store Metadata store Model registry Code repository The MLOps Stack provides one way to think about what tooling you might need. Download template: valohai.com/mlops
Data analysis ML pipeline Model serving Model monitoring Experimentation Feature store Metadata store Model registry Code repository Considerations: Self-Managed Tools vs Managed Tools Point Solutions vs End-to-End Platforms Download template: valohai.com/mlops
The MLOps tools are aplenty… but so are the requirements. Self-driving car vs recommendation engine Citizen data scientists vs seasoned engineers Healthcare & fintech vs mobile gaming Bootstrapping vs well-funded Use Case Team & skills Regulation Other resources
MLOps should fix the development loop. Process Tooling
Check out more reading materials at valohai.com/mlops

More Related Content

What is MLOps

  • 2. What ML development should look like
  • 3. Design 10% Model Development 90% Glue Coding Complex & unstructured handover Operations Time to Market What it actually looks like
  • 5. Why? Because most of your ML system is not about the ML code. Data Verification Serving Infrastructure Configuration Data Collection Feature Extraction Process Management Tools Analysis Tools Monitoring Machine Resource Management ML Code Hidden Technical Debt in Machine Learning Systems (2015): https://papers.nips.cc/paper/2015/file/86df7dcfd896fcaf2674f757a2463eba-Paper.pdf Needed to materialize value
  • 6. The goal of MLOps is to reduce technical friction to get the model from an idea into production in the shortest possible time with as little risk as possible.
  • 7. MLOps is not a single tool or platform. Tooling
  • 8. MLOps is about agreeing to do ML the right way and then supporting it. Process Tooling
  • 9. Process A few shared principles will take you a long way… ML should be collaborative. ML should be reproducible. ML should be continuous. ML should be tested & monitored.
  • 10. Tooling And tooling will help implement your process. ML should be collaborative. ML should be reproducible. ML should be continuous. ML should be tested & monitored. Shared Infrastructure Versioning for Code, Data and Metadata Machine Learning Pipelines Model Deployment and Monitoring
  • 11. Time Risk Short-term thinking Long-term thinking Do it fast Do it safe Time Do it right So it’ll be faster next time Risk Avoid the “It’s too early for MLOps” trap.
  • 12. Data analysis ML pipeline Model serving Model monitoring Experimentation Feature store Metadata store Model registry Code repository The MLOps Stack provides one way to think about what tooling you might need. Download template: valohai.com/mlops
  • 13. Data analysis ML pipeline Model serving Model monitoring Experimentation Feature store Metadata store Model registry Code repository Considerations: Self-Managed Tools vs Managed Tools Point Solutions vs End-to-End Platforms Download template: valohai.com/mlops
  • 14. The MLOps tools are aplenty… but so are the requirements. Self-driving car vs recommendation engine Citizen data scientists vs seasoned engineers Healthcare & fintech vs mobile gaming Bootstrapping vs well-funded Use Case Team & skills Regulation Other resources
  • 15. MLOps should fix the development loop. Process Tooling
  • 16. Check out more reading materials at valohai.com/mlops