![© All Rights Reserved.
21
Some references used to create this presentation
References
Resources
• Melio’s cookiecutter-fastapi (forked from arthurhenrique/cookiecutter-fastapi)
• https://github.com/melio-consulting/cookiecutter-fastapi
• ml-ops.org
• Beyond Jupyter Notebooks: MLOps Environment Setup & First Deployment
• https://www.youtube.com/watch?v=4pkzY95Otm4
• MLOps Stack Canvas:
• https://miro.com/miroverse/mlops-stack-canvas/
• MLOps Is a Mess But That's to be Expected:
• https://www.mihaileric.com/posts/mlops-is-a-mess/
• MLOps at a Reasonable Scale [The Ultimate Guide]:
• https://neptune.ai/blog/mlops-at-reasonable-scale
• Metadata Storage and Management:
• https://mlops.community/learn/metadata-storage-and-management/](https://image.slidesharecdn.com/indabax-mlops-presentation-export-230818082525-30fae282/85/Building-an-MLOps-Stack-for-Companies-at-Reasonable-Scale-15-320.jpg)
Building an MLOps Stack for Companies at Reasonable Scale
- 1. Building an MLOps Stack for Companies at Reasonable Scale ——————————————————————————————————————
- 2. © All Rights Reserved. Is MLOps a Luxury reserved for AI-first enterprises? Is MLOps a Luxury? Intro 85% 11ppl 95 $ k ML projects don’t deliver value Machine Learning Operations (MLOps): Overview, Definition, and Architecture . https://arxiv.org/abs/ 2205.02302 to support an end- to-end ML workflow phData “What is the Cost to Deploy and Maintain a Machine Learning Model?" - https://www.phdata.io/blog/what-is-the-cost-to- deploy-and-maintain-a-machine-learning-model/ to deploy & maintain one ML model Operationalizing Machine Learning: An Interview Study - https://arxiv.org/abs/2209.09125
- 3. © All Rights Reserved. Majority of the companies only need MLOps at reasonable scale Is MLOps a Luxury? Intro Adapted from: MLOps Is a Mess But That’s to be Expected
- 4. © All Rights Reserved. MLOps Ship reliable ML faster - Principles over Technologies - Conventions over Configurations
- 5. © All Rights Reserved. 01 Ship Reliable ML Faster 02 Principle over Technology 03 Convention over Configuration 04 A reasonable MLOps stack 05 Collab to GCP endpoint Demo Practical MLOps
- 6. © All Rights Reserved. Machine Learning + Development + Operations MLOps Principles over Technology Dev Ops Code Infra ML Data Model Data Model Code Adapted from ml-ops.org
- 7. © All Rights Reserved. Technology changes, but good design principles rarely do MLOps Tooling Landscape Principles over Technology
- 8. © All Rights Reserved. Principles that will stand the test of time 7 MLOps Principles Principles over Technology Compliance Reproducibility Versioning Testing Iterative Development Security Monitoring Automation Continuous Deployment Adapted from ml-ops.org
- 9. © All Rights Reserved. Deciding MLOps stack to ship reliable ML faster Decision Framework Conventions over Configurations Tools to Choose Fits 80% of my use case? Run POC Yes Add to Stack Yes Ignore No No Critical operation in Business? Yes No Reversible decision? Expensive? Yes
- 10. © All Rights Reserved. Individual One Team (< 5 DS) Multiple Teams (> 10 DS) Infra / Compute Local / Google Collab AWS Cloud Native Source Control GitHub GitHub GitHub Data Analysis Notebook on Collab Notebook on JupyterHub Notebook on JupyterHub Testing Pytest Pytest Pytest + Others MLOps Stack Level 1: Foundation at reasonable scale * Package Manager, Containerisation, CLI are foundational items and assumed to be present
- 11. © All Rights Reserved. Individual One Team (< 5 DS) Multiple Teams (> 10 DS) Infra / Compute Local / Google Collab AWS Cloud Native Source Control GitHub GitHub GitHub Data ingestion Reading CSVs dbt / Snowflake dbt / Snowflake Data Analysis Notebook on Collab Notebook on JupyterHub Notebook on JupyterHub Experimentation (with HP / NII) Ploomber / Spreadsheets MLFlow + Ray Tune Kubeflow Testing Pytest Pytest Pytest + Others Data Versioning - dvt / Pachyderm dvt / Pachyderm Pipeline Orchestration Cron / Bash Scripts Airflow Kubeflow Pipeline MLOps Stack Level 2: Basic at reasonable scale * Package Manager, Containerisation, CLI are foundational items and assumed to be present
- 12. © All Rights Reserved. Individual One Team (< 5 DS) Multiple Teams (> 10 DS) Infra / Compute Local / Google Collab AWS Cloud Native Source Control GitHub GitHub GitHub / Gitlab Data ingestion Reading CSVs dbt / Snowflake dbt / Snowflake Data Analysis Notebook on Collab Notebook on JupyterHub Notebook on JupyterHub Experimentation Ploomber / Spreadsheets MLFlow + Ray Tune Kubeflow Data Versioning - dvt / Pachyderm dvt / Pachyderm Testing Pytest Pytest Pytest + Others Pipeline Orchestration Cron / Bash Scripts Airflow Kubeflow CI/CD Scripts GitHub Actions Jenkins Model Serving HuggingFace Spaces FastAPI Seldon / kserve Feature / Model Stores - MLFlow Feast + Kubeflow Monitoring Console Logs Grafana + Prometheus Arize AI MLOps Stack Level 3: Advanced at reasonable scale * Package Manager, Containerisation, CLI are foundational items and assumed to be present
- 13. © All Rights Reserved. Individual One Team (< 5 DS) Multiple Teams (> 10 DS) Infra / Compute Source Control Data ingestion Data Analysis Experimentation Data Versioning Testing Pipeline Orchestration CI/CD Model Serving Feature / Model Stores Monitoring MLOps Stack Build your Stack * Package Manager, Containerisation, CLI are foundational items and assumed to be present at reasonable scale Try: https://mymlops.com/
- 14. © All Rights Reserved. Reasonable MLOps Ship reliable ML faster 85% 11ppl 95 $ k without
- 15. © All Rights Reserved. 21 Some references used to create this presentation References Resources • Melio’s cookiecutter-fastapi (forked from arthurhenrique/cookiecutter-fastapi) • https://github.com/melio-consulting/cookiecutter-fastapi • ml-ops.org • Beyond Jupyter Notebooks: MLOps Environment Setup & First Deployment • https://www.youtube.com/watch?v=4pkzY95Otm4 • MLOps Stack Canvas: • https://miro.com/miroverse/mlops-stack-canvas/ • MLOps Is a Mess But That's to be Expected: • https://www.mihaileric.com/posts/mlops-is-a-mess/ • MLOps at a Reasonable Scale [The Ultimate Guide]: • https://neptune.ai/blog/mlops-at-reasonable-scale • Metadata Storage and Management: • https://mlops.community/learn/metadata-storage-and-management/