xDev Asia

ML Platform cho Fashion POD — feature store, training pipeline, model serving, A/B testing, model monitoring, drift detection, MLOps stack (MLflow, registry, experiment tracking).

Kiến trúc Hệ thống Fashion Design & Print-on-Demand — Từ Domain Analysis đến Production

Phần 6: Data Platform & Analytics

xdev.asia

1. ML Platform Overview

Raw Data -> Feature Pipeline -> Feature Store
                     │
                     ├-> Training Jobs -> Model Registry
                     │
                     └-> Online Features -> Model Serving -> Predictions

2. Feature Store Design

interface FeatureDefinition {
  name: string;
  entity: 'user' | 'product' | 'shop';
  type: 'float' | 'int' | 'string' | 'vector';
  source: string;
  ttlHours?: number;
}

const features: FeatureDefinition[] = [
  { name: 'user_30d_click_count', entity: 'user', type: 'int', source: 'events' },
  { name: 'product_ctr_7d', entity: 'product', type: 'float', source: 'analytics' },
  { name: 'product_clip_embedding', entity: 'product', type: 'vector', source: 'ai' },
  { name: 'shop_return_rate_30d', entity: 'shop', type: 'float', source: 'orders' },
];

Offline store: training datasets
Online store: low-latency inference features
Point-in-time correctness: tránh data leakage

3. Training Pipeline

Schedule trigger (daily/weekly)
  -> Build training dataset
  -> Train model
  -> Evaluate metrics
  -> Register candidate model
  -> Optional shadow deployment

class TrainingOrchestrator {
  async run(job: TrainingJob) {
    const dataset = await this.datasetBuilder.build(job.featureSet, job.timeWindow);
    const model = await this.trainer.train(job.algorithm, dataset);
    const metrics = await this.evaluator.evaluate(model, dataset.validation);

    await this.mlflow.logRun({ job, metrics });

    if (metrics.auc >= job.minAuc && metrics.calibrationError <= job.maxCalibrationError) {
      await this.registry.register(model, metrics);
    }
  }
}

4. Model Serving

Pattern	Khi dùng
Online inference	Recommendation, personalization thời gian thực
Batch inference	Nightly ranking precompute, trend prediction
Streaming inference	Fraud/risk scoring theo event

interface PredictionRequest {
  modelName: string;
  entityId: string;
  features: Record<string, unknown>;
}

class ModelServingGateway {
  async predict(req: PredictionRequest) {
    const onlineFeatures = await this.featureStore.getOnline(req.entityId);
    const merged = { ...onlineFeatures, ...req.features };
    return this.runtime.predict(req.modelName, merged);
  }
}

5. A/B Testing Framework

interface Experiment {
  id: string;
  name: string;
  variants: Array<{ name: 'control' | 'treatment'; weight: number }>;
  primaryMetric: 'ctr' | 'conversion' | 'revenue_per_session';
  guardrails: string[];
}

function assignVariant(userId: string, experimentId: string): string {
  const bucket = hash(userId + experimentId) % 100;
  return bucket < 50 ? 'control' : 'treatment';
}

Primary metric rõ ràng trước khi chạy test
Guardrail: latency, error rate, refund rate
Stop criteria: significance + practical impact

6. Monitoring & Drift Detection

Monitors:
- Data drift: PSI / KS distance
- Prediction drift: distribution shift
- Performance drift: CTR/conversion decay
- Operational: latency/error/timeout

if (psi(featureDistTrain, featureDistLive) > 0.2) {
  alert('Feature drift high');
  triggerRetraining('recommendation_model');
}

7. MLOps Governance

Model registry with versioning + approval workflow
Experiment tracking (MLflow)
Reproducible training (code + data snapshot)
Rollback strategy khi model degrade

8. Tổng kết

Feature store là trung tâm để đồng bộ training và serving
Training pipeline cần tiêu chí chất lượng trước khi promote model
A/B testing là cơ chế ra quyết định production an toàn
Drift monitoring giúp phát hiện suy giảm model sớm
MLOps governance đảm bảo ML có thể vận hành bền vững

Bài 21: ML Pipeline & Feature Store — Training, Serving & A/B Testing