# Training Service (ML Model Management)

## Overview

The **Training Service** is the machine learning pipeline engine of Bakery-IA, responsible for training, versioning, and managing Prophet forecasting models. It orchestrates the entire ML workflow from data collection to model deployment, providing real-time progress updates via WebSocket and ensuring bakeries always have the most accurate prediction models. This service enables continuous learning and model improvement without requiring data science expertise.

## Key Features

### Automated ML Pipeline
- **One-Click Model Training** - Train models for all products with a single API call
- **Background Job Processing** - Asynchronous training with job queue management
- **Multi-Product Training** - Process multiple products in parallel
- **Progress Tracking** - Real-time WebSocket updates on training status
- **Automatic Model Versioning** - Track all model versions with performance metrics
- **Model Artifact Storage** - Persist trained models for fast prediction loading

### Training Job Management
- **Job Queue** - FIFO queue for training requests
- **Job Status Tracking** - Monitor pending, running, completed, and failed jobs
- **Concurrent Job Control** - Limit parallel training jobs to prevent resource exhaustion
- **Timeout Handling** - Automatic job termination after maximum duration
- **Error Recovery** - Detailed error messages and retry capabilities
- **Job History** - Complete audit trail of all training executions

### Model Performance Tracking
- **Accuracy Metrics** - MAE, RMSE, R², MAPE for each trained model
- **Historical Comparison** - Compare current vs. previous model performance
- **Per-Product Analytics** - Track which products have the best forecast accuracy
- **Training Duration Tracking** - Monitor training performance and optimization
- **Model Selection** - Automatically deploy best-performing models

### Real-Time Communication
- **WebSocket Live Updates** - Real-time progress percentage and status messages
- **Training Logs** - Detailed step-by-step execution logs
- **Completion Notifications** - RabbitMQ events for training completion
- **Error Alerts** - Immediate notification of training failures

### Feature Engineering
- **Historical Data Aggregation** - Collect sales data for model training
- **External Data Integration** - Fetch weather, traffic, holiday data
- **POI Feature Integration** - Merge location-based POI features into training data
- **Feature Extraction** - Generate 30+ temporal, contextual, and location-based features
- **Data Validation** - Ensure minimum data requirements before training
- **Outlier Detection** - Filter anomalous data points

## Technical Capabilities

### ML Training Pipeline

```python
# Training workflow
async def train_model_pipeline(tenant_id: str, product_id: str):
    """Complete ML training pipeline"""

    # Step 1: Data Collection
    sales_data = await fetch_historical_sales(tenant_id, product_id)
    if len(sales_data) < MIN_TRAINING_DAYS:
        raise InsufficientDataError(f"Need {MIN_TRAINING_DAYS}+ days of data")

    # Step 2: Feature Engineering
    features = engineer_features(sales_data)
    weather_data = await fetch_weather_data(tenant_id)
    traffic_data = await fetch_traffic_data(tenant_id)
    holiday_data = await fetch_holiday_calendar()
    poi_features = await fetch_poi_features(tenant_id)  # NEW: Location context

    # Merge POI features into training dataframe
    features = merge_poi_features(features, poi_features)

    # Step 3: Prophet Model Training
    model = Prophet(
        seasonality_mode='additive',
        daily_seasonality=True,
        weekly_seasonality=True,
        yearly_seasonality=True,
    )
    model.add_country_holidays(country_name='ES')
    model.fit(features)

    # Step 4: Model Validation
    metrics = calculate_performance_metrics(model, sales_data)

    # Step 5: Model Storage
    model_path = save_model_artifact(model, tenant_id, product_id)

    # Step 6: Model Registration
    await register_model_in_database(model_path, metrics)

    # Step 7: Notification
    await publish_training_complete_event(tenant_id, product_id, metrics)

    return model, metrics
```

### WebSocket Progress Updates

```python
# Real-time progress broadcasting
async def broadcast_training_progress(job_id: str, progress: dict):
    """Send progress update to connected clients"""

    message = {
        "type": "training_progress",
        "job_id": job_id,
        "progress": {
            "percentage": progress["percentage"],          # 0-100
            "current_step": progress["step"],              # Step description
            "products_completed": progress["completed"],
            "products_total": progress["total"],
            "estimated_time_remaining": progress["eta"],   # Seconds
            "started_at": progress["start_time"]
        },
        "timestamp": datetime.utcnow().isoformat()
    }

    await websocket_manager.broadcast(job_id, message)
```

### Model Artifact Management

```python
# Model storage and retrieval
import joblib
from pathlib import Path

# Save trained model
def save_model_artifact(model: Prophet, tenant_id: str, product_id: str) -> str:
    """Serialize and store model"""
    model_dir = Path(f"/models/{tenant_id}/{product_id}")
    model_dir.mkdir(parents=True, exist_ok=True)

    version = datetime.utcnow().strftime("%Y%m%d_%H%M%S")
    model_path = model_dir / f"model_v{version}.pkl"

    joblib.dump(model, model_path)
    return str(model_path)

# Load trained model
def load_model_artifact(model_path: str) -> Prophet:
    """Load serialized model"""
    return joblib.load(model_path)
```

### Performance Metrics Calculation

```python
from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score
import numpy as np

def calculate_performance_metrics(model: Prophet, actual_data: pd.DataFrame) -> dict:
    """Calculate comprehensive model performance metrics"""

    # Make predictions on validation set
    predictions = model.predict(actual_data)

    # Calculate metrics
    mae = mean_absolute_error(actual_data['y'], predictions['yhat'])
    rmse = np.sqrt(mean_squared_error(actual_data['y'], predictions['yhat']))
    r2 = r2_score(actual_data['y'], predictions['yhat'])
    mape = np.mean(np.abs((actual_data['y'] - predictions['yhat']) / actual_data['y'])) * 100

    return {
        "mae": float(mae),           # Mean Absolute Error
        "rmse": float(rmse),         # Root Mean Square Error
        "r2_score": float(r2),       # R-squared
        "mape": float(mape),         # Mean Absolute Percentage Error
        "accuracy": float(100 - mape) if mape < 100 else 0.0
    }
```

## Business Value

### For Bakery Owners
- **Continuous Improvement** - Models automatically improve with more data
- **No ML Expertise Required** - One-click training, no data science skills needed
- **Always Up-to-Date** - Weekly automatic retraining keeps models accurate
- **Transparent Performance** - Clear accuracy metrics show forecast reliability
- **Cost Savings** - Automated ML pipeline eliminates need for data scientists

### For Operations Managers
- **Model Version Control** - Track and compare model versions over time
- **Performance Monitoring** - Identify products with poor forecast accuracy
- **Training Scheduling** - Schedule retraining during low-traffic hours
- **Resource Management** - Control concurrent training jobs to prevent overload

### For Platform Operations
- **Scalable ML Pipeline** - Train models for thousands of products
- **Background Processing** - Non-blocking training jobs
- **Error Handling** - Robust error recovery and retry mechanisms
- **Cost Optimization** - Efficient model storage and caching

## Technology Stack

- **Framework**: FastAPI (Python 3.11+) - Async web framework with WebSocket support
- **Database**: PostgreSQL 17 - Training logs, model metadata, job queue
- **ML Library**: Prophet (fbprophet) - Time series forecasting
- **Model Storage**: Joblib - Model serialization
- **File System**: Persistent volumes - Model artifact storage
- **WebSocket**: FastAPI WebSocket - Real-time progress updates
- **Messaging**: RabbitMQ 4.1 - Training completion events
- **ORM**: SQLAlchemy 2.0 (async) - Database abstraction
- **Data Processing**: Pandas, NumPy - Data manipulation
- **Logging**: Structlog - Structured JSON logging
- **Metrics**: Prometheus Client - Custom metrics

## API Endpoints (Key Routes)

### Training Management
- `POST /api/v1/training/start` - Start training job for tenant
- `POST /api/v1/training/start/{product_id}` - Train specific product
- `POST /api/v1/training/stop/{job_id}` - Stop running training job
- `GET /api/v1/training/status/{job_id}` - Get job status and progress
- `GET /api/v1/training/history` - Get training job history
- `DELETE /api/v1/training/jobs/{job_id}` - Delete training job record

### Model Management
- `GET /api/v1/training/models` - List all trained models
- `GET /api/v1/training/models/{model_id}` - Get specific model details
- `GET /api/v1/training/models/{model_id}/metrics` - Get model performance metrics
- `GET /api/v1/training/models/latest/{product_id}` - Get latest model for product
- `POST /api/v1/training/models/{model_id}/deploy` - Deploy specific model version
- `DELETE /api/v1/training/models/{model_id}` - Delete model artifact

### WebSocket
- `WS /api/v1/training/ws/{job_id}` - Connect to training progress stream

### Analytics
- `GET /api/v1/training/analytics/performance` - Overall training performance
- `GET /api/v1/training/analytics/accuracy` - Model accuracy distribution
- `GET /api/v1/training/analytics/duration` - Training duration statistics

## Database Schema

### Main Tables

**training_job_queue**
```sql
CREATE TABLE training_job_queue (
    id UUID PRIMARY KEY,
    tenant_id UUID NOT NULL,
    job_name VARCHAR(255),
    products_to_train TEXT[],          -- Array of product IDs
    status VARCHAR(50) NOT NULL,        -- pending, running, completed, failed
    priority INTEGER DEFAULT 0,
    progress_percentage INTEGER DEFAULT 0,
    current_step VARCHAR(255),
    products_completed INTEGER DEFAULT 0,
    products_total INTEGER,
    started_at TIMESTAMP,
    completed_at TIMESTAMP,
    estimated_completion TIMESTAMP,
    error_message TEXT,
    retry_count INTEGER DEFAULT 0,
    created_by UUID,
    created_at TIMESTAMP DEFAULT NOW(),
    updated_at TIMESTAMP DEFAULT NOW()
);
```

**trained_models**
```sql
CREATE TABLE trained_models (
    id UUID PRIMARY KEY,
    tenant_id UUID NOT NULL,
    product_id UUID NOT NULL,
    model_version VARCHAR(50) NOT NULL,
    model_path VARCHAR(500) NOT NULL,
    training_job_id UUID REFERENCES training_job_queue(id),
    algorithm VARCHAR(50) DEFAULT 'prophet',
    hyperparameters JSONB,
    training_duration_seconds INTEGER,
    training_data_points INTEGER,
    is_deployed BOOLEAN DEFAULT FALSE,
    deployed_at TIMESTAMP,
    created_at TIMESTAMP DEFAULT NOW(),
    UNIQUE(tenant_id, product_id, model_version)
);
```

**model_performance_metrics**
```sql
CREATE TABLE model_performance_metrics (
    id UUID PRIMARY KEY,
    model_id UUID REFERENCES trained_models(id),
    tenant_id UUID NOT NULL,
    product_id UUID NOT NULL,
    mae DECIMAL(10, 4),                -- Mean Absolute Error
    rmse DECIMAL(10, 4),               -- Root Mean Square Error
    r2_score DECIMAL(10, 6),           -- R-squared
    mape DECIMAL(10, 4),               -- Mean Absolute Percentage Error
    accuracy_percentage DECIMAL(5, 2),
    validation_data_points INTEGER,
    created_at TIMESTAMP DEFAULT NOW()
);
```

**model_training_logs**
```sql
CREATE TABLE model_training_logs (
    id UUID PRIMARY KEY,
    training_job_id UUID REFERENCES training_job_queue(id),
    tenant_id UUID NOT NULL,
    product_id UUID,
    log_level VARCHAR(20),              -- DEBUG, INFO, WARNING, ERROR
    message TEXT,
    step_name VARCHAR(100),
    execution_time_ms INTEGER,
    metadata JSONB,
    created_at TIMESTAMP DEFAULT NOW()
);
```

**model_artifacts** (Metadata only, actual files on disk)
```sql
CREATE TABLE model_artifacts (
    id UUID PRIMARY KEY,
    model_id UUID REFERENCES trained_models(id),
    artifact_type VARCHAR(50),          -- model_file, feature_list, scaler, etc.
    file_path VARCHAR(500),
    file_size_bytes BIGINT,
    checksum VARCHAR(64),               -- SHA-256 hash
    created_at TIMESTAMP DEFAULT NOW()
);
```

## Events & Messaging

### Published Events (RabbitMQ)

**Exchange**: `training`
**Routing Key**: `training.completed`

**Training Completed Event**
```json
{
    "event_type": "training_completed",
    "tenant_id": "uuid",
    "job_id": "uuid",
    "job_name": "Weekly retraining - All products",
    "status": "completed",
    "results": {
        "successful_trainings": 25,
        "failed_trainings": 2,
        "total_products": 27,
        "models_created": [
            {
                "product_id": "uuid",
                "product_name": "Baguette",
                "model_version": "20251106_143022",
                "accuracy": 82.5,
                "mae": 12.3,
                "rmse": 18.7,
                "r2_score": 0.78
            }
        ],
        "average_accuracy": 79.8,
        "training_duration_seconds": 342
    },
    "started_at": "2025-11-06T14:25:00Z",
    "completed_at": "2025-11-06T14:30:42Z",
    "timestamp": "2025-11-06T14:30:42Z"
}
```

**Training Failed Event**
```json
{
    "event_type": "training_failed",
    "tenant_id": "uuid",
    "job_id": "uuid",
    "product_id": "uuid",
    "product_name": "Croissant",
    "error_type": "InsufficientDataError",
    "error_message": "Product requires minimum 30 days of sales data. Currently: 15 days.",
    "recommended_action": "Collect more sales data before retraining",
    "severity": "medium",
    "timestamp": "2025-11-06T14:28:15Z"
}
```

### Consumed Events
- **From Orchestrator**: Scheduled training triggers
- **From Sales**: New sales data imported (triggers retraining)

## Custom Metrics (Prometheus)

```python
# Training job metrics
training_jobs_total = Counter(
    'training_jobs_total',
    'Total training jobs started',
    ['tenant_id', 'status']  # completed, failed, cancelled
)

training_duration_seconds = Histogram(
    'training_duration_seconds',
    'Training job duration',
    ['tenant_id'],
    buckets=[10, 30, 60, 120, 300, 600, 1800, 3600]  # seconds
)

models_trained_total = Counter(
    'models_trained_total',
    'Total models successfully trained',
    ['tenant_id', 'product_category']
)

# Model performance metrics
model_accuracy_distribution = Histogram(
    'model_accuracy_percentage',
    'Distribution of model accuracy scores',
    ['tenant_id'],
    buckets=[50, 60, 70, 75, 80, 85, 90, 95, 100]  # percentage
)

model_mae_distribution = Histogram(
    'model_mae',
    'Distribution of Mean Absolute Error',
    ['tenant_id'],
    buckets=[1, 5, 10, 20, 30, 50, 100]  # units
)

# WebSocket metrics
websocket_connections_total = Gauge(
    'training_websocket_connections',
    'Active WebSocket connections',
    ['tenant_id']
)

websocket_messages_sent = Counter(
    'training_websocket_messages_total',
    'Total WebSocket messages sent',
    ['tenant_id', 'message_type']
)
```

## Configuration

### Environment Variables

**Service Configuration:**
- `PORT` - Service port (default: 8004)
- `DATABASE_URL` - PostgreSQL connection string
- `RABBITMQ_URL` - RabbitMQ connection string
- `MODEL_STORAGE_PATH` - Path for model artifacts (default: /models)

**Training Configuration:**
- `MAX_CONCURRENT_JOBS` - Maximum parallel training jobs (default: 3)
- `MAX_TRAINING_TIME_MINUTES` - Job timeout (default: 30)
- `MIN_TRAINING_DATA_DAYS` - Minimum history required (default: 30)
- `ENABLE_AUTO_DEPLOYMENT` - Auto-deploy after training (default: true)

**Prophet Configuration:**
- `PROPHET_DAILY_SEASONALITY` - Enable daily patterns (default: true)
- `PROPHET_WEEKLY_SEASONALITY` - Enable weekly patterns (default: true)
- `PROPHET_YEARLY_SEASONALITY` - Enable yearly patterns (default: true)
- `PROPHET_INTERVAL_WIDTH` - Confidence interval (default: 0.95)
- `PROPHET_CHANGEPOINT_PRIOR_SCALE` - Trend flexibility (default: 0.05)

**WebSocket Configuration:**
- `WEBSOCKET_HEARTBEAT_INTERVAL` - Ping interval seconds (default: 30)
- `WEBSOCKET_MAX_CONNECTIONS` - Max connections per tenant (default: 10)
- `WEBSOCKET_MESSAGE_QUEUE_SIZE` - Message buffer size (default: 100)

**Storage Configuration:**
- `MODEL_RETENTION_DAYS` - Days to keep old models (default: 90)
- `MAX_MODEL_VERSIONS_PER_PRODUCT` - Version limit (default: 10)
- `ENABLE_MODEL_COMPRESSION` - Compress model files (default: true)

## Development Setup

### Prerequisites
- Python 3.11+
- PostgreSQL 17
- RabbitMQ 4.1
- Persistent storage for model artifacts

### Local Development
```bash
# Create virtual environment
cd services/training
python -m venv venv
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Set environment variables
export DATABASE_URL=postgresql://user:pass@localhost:5432/training
export RABBITMQ_URL=amqp://guest:guest@localhost:5672/
export MODEL_STORAGE_PATH=/tmp/models

# Create model storage directory
mkdir -p /tmp/models

# Run database migrations
alembic upgrade head

# Run the service
python main.py
```

### Testing
```bash
# Unit tests
pytest tests/unit/ -v

# Integration tests (requires services)
pytest tests/integration/ -v

# WebSocket tests
pytest tests/websocket/ -v

# Test with coverage
pytest --cov=app tests/ --cov-report=html
```

### WebSocket Testing
```python
# Test WebSocket connection
import asyncio
import websockets
import json

async def test_training_progress():
    uri = "ws://localhost:8004/api/v1/training/ws/job-id-here"
    async with websockets.connect(uri) as websocket:
        while True:
            message = await websocket.recv()
            data = json.loads(message)
            print(f"Progress: {data['progress']['percentage']}%")
            print(f"Step: {data['progress']['current_step']}")

            if data['type'] == 'training_completed':
                print("Training finished!")
                break

asyncio.run(test_training_progress())
```

## POI Feature Integration

### How POI Features Enhance Training

The Training Service integrates location-based POI features from the External Service to improve forecast accuracy:

**POI Features Included:**
- `school_density` - Number of schools within 1km radius
- `office_density` - Number of offices and business centers nearby
- `residential_density` - Residential area proximity
- `transport_hub_proximity` - Distance to metro, bus, train stations
- `commercial_zone_score` - Commercial activity in the area
- `restaurant_density` - Nearby restaurants and cafes
- `competitor_proximity` - Distance to competing bakeries
- And 11+ more location-based features

**Integration Process:**
1. **Fetch POI Context** - Retrieve tenant's POI features from External Service (`/poi-context/{tenant_id}`)
2. **Extract ML Features** - Parse `ml_features` JSON object from POI context
3. **Merge with Training Data** - Add POI features as additional columns in training dataframe
4. **Prophet Training** - Include POI features as regressors in Prophet model
5. **Feature Importance** - Track which POI features most impact predictions

**Example POI Feature Integration:**
```python
from app.ml.poi_feature_integrator import POIFeatureIntegrator

# Initialize POI integrator
poi_integrator = POIFeatureIntegrator(external_service_url)

# Fetch and merge POI features
poi_features = await poi_integrator.fetch_poi_features(tenant_id)
training_df = poi_integrator.merge_poi_features(training_df, poi_features)

# POI features now available as columns:
# training_df['school_density'], training_df['office_density'], etc.

# Add POI features as Prophet regressors
for feature_name in poi_features.keys():
    prophet_model.add_regressor(feature_name)
```

**Endpoint Used:**
- Via shared client: `/api/v1/tenants/{tenant_id}/external/poi-context` (routed through API Gateway)

## Integration Points

### Dependencies (Services Called)
- **Sales Service** - Fetch historical sales data for training
- **External Service** - Fetch weather, traffic, holiday, and POI feature data
- **PostgreSQL** - Store job queue, models, metrics, logs
- **RabbitMQ** - Publish training completion events
- **File System** - Store model artifacts

### Dependents (Services That Call This)
- **Forecasting Service** - Load trained models for predictions
- **Orchestrator Service** - Trigger scheduled training jobs
- **Frontend Dashboard** - Display training progress and model metrics
- **AI Insights Service** - Analyze model performance patterns

## Security Measures

### Data Protection
- **Tenant Isolation** - All training jobs scoped to tenant_id
- **Model Access Control** - Only tenant can access their models
- **Input Validation** - Validate all training parameters
- **Rate Limiting** - Prevent training job spam

### Model Security
- **Model Checksums** - SHA-256 hash verification for artifacts
- **Version Control** - Track all model versions with audit trail
- **Access Logging** - Log all model access and deployment
- **Secure Storage** - Model files stored with restricted permissions

### WebSocket Security
- **JWT Authentication** - Authenticate WebSocket connections
- **Connection Limits** - Max connections per tenant
- **Message Validation** - Validate all WebSocket messages
- **Heartbeat Monitoring** - Detect and close stale connections

## Performance Optimization

### Training Performance
1. **Parallel Processing** - Train multiple products concurrently
2. **Data Caching** - Cache fetched external data across products
3. **Incremental Training** - Only retrain changed products
4. **Resource Limits** - CPU/memory limits per training job
5. **Priority Queue** - Prioritize important products first

### Storage Optimization
1. **Model Compression** - Compress model artifacts (gzip)
2. **Old Model Cleanup** - Automatic deletion after retention period
3. **Version Limits** - Keep only N most recent versions
4. **Deduplication** - Avoid storing identical models

### WebSocket Optimization
1. **Message Batching** - Batch progress updates (every 2 seconds)
2. **Connection Pooling** - Reuse WebSocket connections
3. **Compression** - Enable WebSocket message compression
4. **Heartbeat** - Keep connections alive efficiently

## Troubleshooting

### Common Issues

**Issue**: Training jobs stuck in "pending" status
- **Cause**: Max concurrent jobs reached or worker process crashed
- **Solution**: Check `MAX_CONCURRENT_JOBS` setting, restart service

**Issue**: WebSocket connection drops during training
- **Cause**: Network timeout or client disconnection
- **Solution**: Implement auto-reconnect logic in client

**Issue**: "Insufficient data" errors for many products
- **Cause**: Products need 30+ days of sales history
- **Solution**: Import more historical sales data or reduce `MIN_TRAINING_DATA_DAYS`

**Issue**: Low model accuracy (<70%)
- **Cause**: Insufficient data, outliers, or changing business patterns
- **Solution**: Clean outliers, add more features, or manually adjust Prophet params

### Debug Mode
```bash
# Enable detailed logging
export LOG_LEVEL=DEBUG
export PROPHET_VERBOSE=1

# Enable training profiling
export ENABLE_PROFILING=1

# Disable concurrent jobs for debugging
export MAX_CONCURRENT_JOBS=1
```

## Competitive Advantages

1. **One-Click ML** - No data science expertise required
2. **Real-Time Visibility** - WebSocket progress updates unique in bakery software
3. **Continuous Learning** - Automatic weekly retraining
4. **Version Control** - Track and compare all model versions
5. **Production-Ready** - Robust error handling and retry mechanisms
6. **Scalable** - Train models for thousands of products
7. **Spanish Market** - Optimized for Spanish bakery patterns and holidays

## Future Enhancements

- **Hyperparameter Tuning** - Automatic optimization of Prophet parameters
- **A/B Testing** - Deploy multiple models and compare performance
- **Distributed Training** - Scale across multiple machines
- **GPU Acceleration** - Use GPUs for deep learning models
- **AutoML** - Automatic algorithm selection (Prophet vs LSTM vs ARIMA)
- **Model Explainability** - SHAP values to explain predictions
- **Custom Algorithms** - Support for user-provided ML models
- **Transfer Learning** - Use pre-trained models from similar bakeries

---

**For VUE Madrid Business Plan**: The Training Service demonstrates advanced ML engineering capabilities with automated pipeline management and real-time monitoring. The ability to continuously improve forecast accuracy without manual intervention represents significant operational efficiency and competitive advantage. This self-learning system is a key differentiator in the bakery software market and showcases technical innovation suitable for EU technology grants and investor presentations.