The ML Lifecycle

Unlike traditional software with a linear development cycle, ML systems follow a continuous loop. Understanding this lifecycle is key to building effective MLOps practices.

The ML Development Loop

    ┌─────────────────────────────────────────┐
    │                                         │
    ▼                                         │
┌────────┐    ┌────────┐    ┌────────┐    ┌──┴─────┐
│  Data  │───▶│ Train  │───▶│ Deploy │───▶│Monitor │
└────────┘    └────────┘    └────────┘    └────────┘
    ▲                                         │
    │                                         │
    └─────────── Retrain ◄────────────────────┘

Stage Breakdown

Stage	Activities	Key Tools
Data	Collection, validation, versioning	DVC, Great Expectations
Train	Experimentation, model building	MLflow, W&B
Deploy	Packaging, serving, scaling	BentoML, KServe
Monitor	Performance tracking, drift detection	Evidently, Arize
Retrain	Trigger detection, automated pipelines	Kubeflow, Airflow

Data Stage

The foundation of any ML system. Poor data leads to poor models.

# Example: DVC for data versioning
# Initialize DVC in your project
# $ dvc init

# Track a dataset
# $ dvc add data/training_data.csv

# This creates:
# - data/training_data.csv.dvc  (metadata)
# - .gitignore updated to exclude large file

Key activities:

• Data collection and ingestion
• Data validation and quality checks
• Feature engineering
• Data versioning and lineage

Train Stage

Where data becomes models through experimentation.

import mlflow

# Track experiments with MLflow
mlflow.set_experiment("customer-churn")

with mlflow.start_run():
    # Log parameters
    mlflow.log_param("model_type", "random_forest")
    mlflow.log_param("n_estimators", 100)

    # Train model
    model = train_model(X_train, y_train)

    # Log metrics
    mlflow.log_metric("accuracy", evaluate(model, X_test, y_test))

    # Save model
    mlflow.sklearn.log_model(model, "model")

Key activities:

• Hyperparameter tuning
• Model selection
• Experiment tracking
• Model validation

Deploy Stage

Moving models from notebooks to production.

# Example: BentoML service definition
import bentoml

@bentoml.service
class ChurnPredictor:
    def __init__(self):
        self.model = load_model()

    @bentoml.api
    def predict(self, features: dict) -> float:
        return self.model.predict([features])[0]

Key activities:

• Model packaging
• Serving infrastructure
• A/B testing
• Rollback strategies

Monitor Stage

Production models need constant observation.

What to Monitor	Why
Prediction latency	User experience, SLAs
Data drift	Input distribution changes
Model accuracy	Performance degradation
Resource usage	Cost optimization

The Continuous Loop

Unlike "deploy and done" software, ML systems require:

1. Continuous monitoring - Models degrade over time
2. Automatic triggers - Detect when retraining is needed
3. Automated pipelines - Retrain without manual intervention
4. Gradual rollouts - Deploy new models safely

Key insight: The ML lifecycle is not a waterfall—it's a flywheel. Each iteration should improve the system.

Next, we'll explore MLOps maturity levels and what distinguishes advanced organizations. :::