Lesson 1: What is MLOps? — ML Lifecycle & Maturity Levels

MLOps & LLMOps: Bringing AI to Production

Part 1: MLOps Foundations

xdev.asia

Introduction

87% of ML models never go to production. The problem isn't that the model is bad — it's that there is no process for getting the model from the notebook to the real world. MLOps solves this.

🎯 MLOps = Machine Learning + DevOps + Data Engineering

---

1. ML Lifecycle — ML Project life cycle

┌─────────────────────────────────────────────────────────┐
│                    ML LIFECYCLE                          │
│                                                         │
│  1. Problem     2. Data        3. Feature               │
│     Definition     Collection     Engineering           │
│         │              │              │                  │
│         ▼              ▼              ▼                  │
│  4. Model       5. Training    6. Evaluation            │
│     Selection      & Tuning       & Validation          │
│         │              │              │                  │
│         ▼              ▼              ▼                  │
│  7. Deployment  8. Monitoring  9. Retraining            │
│     & Serving      & Alerts       & Updates             │
│         │              │              │                  │
│         └──────────────┴──────────────┘                  │
│                   (Continuous Loop)                      │
└─────────────────────────────────────────────────────────┘

Each step has its own problems:

Phase	Common problems
Data	Data changes, schema drift, quality issues
Training	Not reproducible, lost track experiments
Evaluation	Offline metrics ≠ online performance
Deployment	"Runs on my computer" but fails on the server
Monitoring	Model decay, don't know when to retrain

---

2. DevOps vs MLOps

DevOps (Software):
  Code → Build → Test → Deploy → Monitor
  ✅ Deterministic (cùng code → cùng output)

MLOps (Machine Learning):
  Data + Code + Config → Train → Evaluate → Deploy → Monitor
  ❌ Non-deterministic (cùng code, khác data → khác model)
  ❌ Data dependency (model phụ thuộc vào data quality)
  ❌ Model decay (model giảm chất lượng theo thời gian)

Key differences:

	DevOps	MLOps
Artifact	Binary / Container	Model + Data + Config
Testing	Unit tests, integration	+ Data validation, model validation
CI/CD	Code changes	+ Data changes, model retraining
Monitoring	Uptime, latency	+ Data drift, model performance
Versioning	Code (Git)	+ Data + Model + Pipeline
Reproducibility	Easy	Very difficult (random seeds, GPU, ...)

---

3. MLOps Maturity Model

Level 0: No MLOps (Manual)

Đặc điểm:
  ❌ Jupyter Notebook → Manual deploy
  ❌ Không track experiments
  ❌ Không monitoring
  ❌ Retrain = "ai đó nhớ thì làm"

Team:
  1 Data Scientist làm hết

Phù hợp: POC, hackathon

Level 1: DevOps but not yet MLOps

Đặc điểm:
  ✅ Code trên Git
  ✅ CI/CD pipeline
  ✅ Automated testing (unit tests)
  ❌ Chưa track data versions
  ❌ Chưa track experiments
  ❌ Manual retraining

Team:
  DS + ML Engineer

Phù hợp: Startup giai đoạn đầu

Level 2: ML Pipeline Automation

Đặc điểm:
  ✅ Automated training pipeline
  ✅ Experiment tracking (MLflow)
  ✅ Data versioning (DVC)
  ✅ Model registry
  ✅ Feature store
  ⚠️ Manual trigger retraining

Team:
  DS + ML Engineer + Data Engineer

Phù hợp: Công ty có 5-10 ML models

Level 3: CI/CD for ML

Đặc điểm:
  ✅ Automated retraining (trigger by data/schedule)
  ✅ A/B testing, canary deployment
  ✅ Model validation pipeline
  ✅ Monitoring + alerting
  ✅ Feature store shared

Team:
  DS + ML Engineer + Data Engineer + ML Platform

Phù hợp: Công ty scale (>10 models)

Level 4: Full MLOps

Đặc điểm:
  ✅ Self-healing pipelines
  ✅ Auto-retrain on drift detection
  ✅ Multi-model management
  ✅ Cost optimization
  ✅ Governance & compliance

Team:
  Full ML Platform team

Phù hợp: Big Tech, AI-first companies

---

4. Technical Debt in ML

Google's "Hidden Technical Debt in ML Systems" (NeurIPS 2015):

┌────────────────────────────────────────────┐
│              ML System                      │
│  ┌────────────────────────────────────┐    │
│  │         ML Code (~5%)              │    │
│  └────────────────────────────────────┘    │
│  ┌────┬─────┬──────┬──────┬────┬──────┐   │
│  │Data│Data │Feat. │Config│Serv│Monit.│   │
│  │Col.│Veri.│Extr. │     │ing │oring │   │
│  └────┴─────┴──────┴──────┴────┴──────┘   │
│              (~95% non-ML code)            │
└────────────────────────────────────────────┘

ML Code chỉ chiếm ~5% tổng hệ thống!

Types of technical debt:

# 1. Data Dependency Debt
# Input data thay đổi → model hỏng
# VD: Feature từ API bên thứ 3 bị đổi format

# 2. Configuration Debt
# Hyperparams, feature flags, thresholds... không tracked
# VD: Ai đổi threshold từ 0.5 → 0.7? Khi nào?

# 3. Pipeline Debt
# Glue code nối các bước → fragile
# VD: Script bash + cron job + manual copy file

# 4. Reproducibility Debt
# Không thể reproduce kết quả cũ
# VD: "Model v2 tốt hơn v1" — nhưng không reproduce được v1

---

5. MLOps Tools Ecosystem

┌─────────────────────────────────────────────────┐
│                 MLOps Stack                      │
├────────────┬────────────────────────────────────┤
│ Layer      │ Tools                               │
├────────────┼────────────────────────────────────┤
│ Experiment │ MLflow, W&B, Neptune, CometML      │
│ Tracking   │                                     │
├────────────┼────────────────────────────────────┤
│ Data Vers. │ DVC, LakeFS, Delta Lake            │
├────────────┼────────────────────────────────────┤
│ Feature    │ Feast, Tecton, Hopsworks           │
│ Store      │                                     │
├────────────┼────────────────────────────────────┤
│ Model Reg. │ MLflow, Vertex AI, SageMaker       │
├────────────┼────────────────────────────────────┤
│ Orchest.   │ Kubeflow, Airflow, Prefect         │
├────────────┼────────────────────────────────────┤
│ Serving    │ TorchServe, Triton, TFServing, BentoML │
├────────────┼────────────────────────────────────┤
│ Monitoring │ Evidently, Arize, WhyLabs          │
├────────────┼────────────────────────────────────┤
│ Infra      │ Docker, K8s, Terraform             │
└────────────┴────────────────────────────────────┘

Tool selected according to team size:

Team Size	Recommendation
1-3 people	MLflow + DVC + Docker
3-10 people	+ Airflow + Feast + Evidently
10+ people	Full platform: Kubeflow / Vertex AI / SageMaker
Enterprise	Commercial: Databricks, Dataiku, Domino

---

6. Hands-on: Setup MLOps Project

"""Setup cấu trúc project MLOps chuẩn"""

# Project structure
project_structure = """
my-ml-project/
├── data/
│   ├── raw/              # Data gốc (never modify)
│   ├── processed/        # Data sau preprocessing
│   └── features/         # Feature store output
├── notebooks/            # EDA, prototyping
├── src/
│   ├── data/             # Data processing code
│   ├── features/         # Feature engineering
│   ├── models/           # Model training code
│   ├── serving/          # Inference server
│   └── monitoring/       # Monitoring code
├── configs/
│   ├── training.yaml     # Training hyperparameters
│   ├── serving.yaml      # Serving config
│   └── monitoring.yaml   # Alerting rules
├── tests/
│   ├── test_data.py      # Data validation tests
│   ├── test_model.py     # Model validation tests
│   └── test_api.py       # API tests
├── pipelines/
│   ├── training.py       # Training pipeline
│   ├── evaluation.py     # Evaluation pipeline
│   └── deployment.py     # Deployment pipeline
├── Dockerfile
├── docker-compose.yml
├── Makefile              # Common commands
├── dvc.yaml              # DVC pipeline
├── mlflow.yaml           # MLflow config
└── README.md
"""

print(project_structure)

# Makefile — Common commands
.PHONY: setup train evaluate deploy monitor

setup:
	pip install -r requirements.txt
	dvc pull

train:
	python pipelines/training.py --config configs/training.yaml

evaluate:
	python pipelines/evaluation.py --model-version latest

deploy:
	python pipelines/deployment.py --target production

monitor:
	python src/monitoring/check_drift.py

test:
	pytest tests/ -v

lint:
	ruff check src/
	mypy src/

---

Summary

Concepts	Remember
MLOps	ML + DevOps + Data Engineering
ML Lifecycle	Data → Train → Deploy → Monitor → Retrain
Maturity 0-4	Manual → Automated → CI/CD → Full MLOps
Tech Debt	ML code is only ~5%, 95% is infrastructure
Tools	MLflow, DVC, Feast, Kubeflow, Evidently

Exercises

1. Evaluate your team: What maturity level is your team at? List gaps.
2. Project Setup: Create project structure according to the above template. Init git + dvc.
3. Research Tools: Compare 2 experiment tracking tools: MLflow vs Weights & Biases. Choose 1 for the team.

Next article: Experiment Tracking — MLflow & Weights & Biases.