Combinare le pratiche DevOps e le potenzialità di GitHub per il rilascio continuo e condizionato di prodotti di AI su Azure

1. Approcci DevOps per il
Rilascio Continuo di Modelli
ML su Azure
Igor Antonacci – Automation Lover
Luca Milan – IT & DevOps addicted
Combinare le pratiche DevOps e le potenzialità di
GitHub per il rilascio continuo e condizionato
di prodotti di AI su Azure.

2. Platinum Sponsor
Gold Sponsor
Technical Sponsor

3. Agenda
❑Introduzione
❑DevOps & Machine Learning = MLOps
❑Tools e Metodologie
❑Demo Time

4. © Copyright Microsoft Corporation. All rights reserved.
Computer Science
AI
Development
IT
Operations

5. ML: Peculiarità
Matrice
Matematica
Strumenti
Dedicati
Ciclo di
Vita
Modalità
Rilascio
Codice
Dati
Modello

6. ML: Ciclo di Vita
Training
Data
Learning
Process
Learned
Function
Scoring
or
Serving
Get fresh data and training again

7. Machine Learning senza Python

8. DevOps Pipeline 4 ML
Data or Algorithm
updated
Function Code
updated
Infrastructure
updated
ML.NET
Training
Deploy
Model
Pulumi Up
Deploy
Function
Yes
DevOps
Done
No
Build & Test
Triggers
C# Code

9. Scoring/Serving
Micro UI
ML Training
Predictor
Micro UI
Code + Data
Github Pipelines
(Re)Feed ML Pipeline
Solution
Overview

11. IGOR ANTONACCI

12. © Copyright Microsoft Corporation. All rights reserved.
Obiettivo organizzativo
All’interno di un Team di sviluppo, facilitare il dispiegamento di
Modelli di IA attraverso:
• la COOPERAZIONE tra Data Scientists e IT Operations
• L’AUTOMAZIONE del processo di Analisi, Sviluppo, Test e
rilascio di un modello

13. © Copyright Microsoft Corporation. All rights reserved.
MLOps
“With Machine Learning Model Operationalization Management (MLOps),
we want to provide an end-to-end machine learning development process
to design, build and manage reproducible, testable, and evolvable ML-
powered software.”
(cit. ML Ops: Machine Learning Operations (ml-ops.org))

14. Solution Architecture
Prediction Pipeline
v0.1.0-abcdef
Training Model Pipeline
Micro-UI Pipeline

15. Branching model
for Data Scientists
MAIN
experiment / model-refininig_001
experiment / ml-net_release-change_001
Pull Request
✓ Model
(Code)
✓ Test
✓ Dataset
✓ Review
✓ Approve
✓ Merge

16. GitHub Actions to the
rescue
COMMIT
✓ Model (Code)
✓ Test
✓ Dataset
✓ Build
✓ Test
✓ Training
Pull-request
✓ Review
✓ Approve
✓ Merge
Training Model

17. GitHub Environments
✓ on push
✓ on pull-request
✓ …
Job #1 Job #2
Action 1 – Setup .NET Core
Action 2 – Build project
Action 3 – Test project
Action N – …
Action 1 – GitVersion
Action 2 – Create Release
Action 3 – Publish Artifacts
Action N – …
Depends on
Secrets
✓ on push
✓ on pull-request
✓ …
Job #1 Job #2
Action 1 – Setup .NET Core
Action 2 – Build project
Action 3 – Test project
Action N – …
Action 1 – GitVersion
Action 2 – Create Release
Action 3 – Publish Artifacts
Action N – …
Depends on
Environment
Approvers

18. GitHub Environments
COMMIT
✓ Model (Code)
✓ Test
✓ Dataset
✓ Build
✓ Test
✓ Training
✓ Trained Model
✓ Dataset
Pull-request
✓ Review
✓ Approve
✓ Merge
Approvazione
Promozione
✓ Review
✓ Approve
Serving “Environment”
Training Model
Dispatch Message
▪ event_type: model-created
▪ message: |
{
"ML_MODEL_VERSION": vX.Y.Z-xxxx,
"ML_MODEL_URI":
https://github.com/igoran/.../${{
env.VERSION_NAME }}.zip
}

19. GitHub as Artifact
Repository
COMMIT
✓ Model (Code)
✓ Test
✓ Dataset
✓ Build
✓ Test
✓ Training
✓ Trained Model
✓ Dataset
Pull-request
✓ Review
✓ Approve
✓ Merge
Approvazione
Promozione
✓ Review
✓ Approve
Serving “Environment”
Dispatch Message
▪ event_type: model-created
▪ message: |
{
"ML_MODEL_VERSION": vX.Y.Z-xxxx,
"ML_MODEL_URI":
https://github.com/igoran/.../${{
env.VERSION_NAME }}.zip
}
Training Model
0.1.0-c28c8ba

20. GH Actions – Workflow
Dispatch
COMMIT
✓ Model (Code)
✓ Test
✓ Dataset
✓ Build
✓ Test
✓ Training
✓ Trained Model
✓ Dataset
Pull-request
✓ Review
✓ Approve
✓ Merge
Approvazione
Promozione
✓ Review
✓ Approve
Serving “Environment”
Training Model
Dispatch Message
▪ event_type: model-created
▪ message: |
{
"ML_MODEL_VERSION": vX.Y.Z-xxxx,
"ML_MODEL_URI":
https://github.com/igoran/.../${{
env.VERSION_NAME }}.zip
}

21. GH Actions – Workflow
Dispatch
SETUP INFRA
✓ Build
✓ Test
Azure
Deployment
Slots
model-created
✓ Trained Model
✓ Dataset
Download Model
PROD DEV
Predictor

22. MicroUI Pipeline
DEPLOY
✓ Build
✓ Test
FrontEnd Dev

23. Scoring/Serving
Micro UI
ML Training
Predictor
Micro UI
Code + Data
Github Pipelines
(Re)Feed ML Pipeline
Solution
Overview

24. © Copyright Microsoft Corporation. All rights reserved.
PULUMI
▪ Infrastructure-as-Code Framework
▪ Multi-Platform Providers (AWS, Azure, GCP … )
▪ (Multi-)Language oriented
▪ C#, Python, Go …
▪ CLI ed Engine multi piattaforma (pulumi)
▪ NO YAML ☺
▪ Ambienti “sandboxed” →Stacks

25. DEMO

27. Platinum Sponsor
Gold Sponsor
Technical Sponsor