Materi 13 · Applied AI

MLOps & Deployment

Model di Jupyter Notebook ≠ model production. Pelajari pipeline ML end-to-end: data versioning, model registry, deployment, monitoring, dan continuous training.

⏱ 30 Menit🎯 Advanced📚 Module 13/15

1. Apa Itu MLOps?

MLOps = praktik DevOps untuk machine learning. Mencakup automation, version control, monitoring, dan governance dari ML model di production.

"Hidden Technical Debt in ML Systems"

Paper Google 2015: ML code hanya ~5% dari ML system. 95% sisanya: data pipeline, feature engineering, deployment, monitoring, infrastructure. MLOps mengelola 95% ini.

2. ML Pipeline End-to-End

1Data Collection

2Data Validation

3Feature Engineering

4Model Training

5Model Eval

6Deployment

7Monitoring

3. Tooling Stack Modern

📦

Data Versioning

DVC, LakeFS, Pachyderm — git untuk dataset.

🧪

Experiment Tracking

MLflow, Weights & Biases, Neptune — log run, params, metrics.

🏭

Pipeline Orchestration

Airflow, Kubeflow, Prefect, Dagster.

📚

Model Registry

MLflow Registry, SageMaker Registry — versi model.

🚀

Serving

TorchServe, Triton, BentoML, FastAPI, KServe.

📊

Monitoring

Evidently, Arize, WhyLabs — drift detection.

4. Deployment Patterns

Pattern	Karakter	Cocok untuk
Batch Prediction	Predict offline, simpan ke DB	Recommendations harian
Real-time API	REST/gRPC endpoint	Fraud detection, search
Streaming	Predict pada event stream	Anomaly detection IoT
Edge / On-device	Model di smartphone/IoT	Privacy, low-latency
Embedded	Tertanam di app/library	Speech recognition

5. Monitoring di Production

3 Tipe Drift yang Wajib Di-Monitor

1. Data Drift: distribusi input feature berubah dari training data. Mis. usia user shift dari 25-35 ke 18-25.
2. Concept Drift: hubungan input-output berubah. Mis. preferensi konsumen pasca-pandemic.
3. Model Performance Drift: accuracy turun seiring waktu. Cek dengan ground truth (jika tersedia).

Statistical tests: KS test, chi-square untuk detect distribution shift.
Performance metric: log accuracy, latency, error rate di production.
Feedback loop: kumpulkan ground truth dari user → re-evaluate model.
Alerting: auto-trigger retraining ketika drift > threshold.

6. CI/CD untuk ML

CI/CD/CT Pipeline CI: unit test, integration test, data validation
CD: deploy model ke staging → production
CT: Continuous Training — auto retrain saat ada
drift, data baru, atau scheduled cadence

7. Deployment Strategy

Shadow Mode: deploy model baru paralel, log prediksinya tanpa mempengaruhi user. Validasi sebelum live.
Canary Release: route 5% traffic ke model baru, naik bertahap kalau metric oke.
A/B Test: bagi 50/50 traffic ke 2 versi model. Bandingkan business metric, bukan hanya accuracy.
Blue-Green: 2 environment paralel. Switch sekaligus saat siap.
Feature Flag: control rollout per user/segment dari config.

8. Studi Kasus

🌟 Real World

Gojek: Platform ML untuk 200+ Model di Production

Gojek punya internal platform ML (codename: "Marvel") yang melayani 200+ model di production: dynamic pricing, ETA prediction, fraud detection, recommendation, dst. Investasi besar di MLOps memungkinkan tim data scientist fokus pada modeling, bukan engineering.

Pelajaran: tanpa platform MLOps, scaling ML di company besar = chaos. Setiap tim reinvent pipeline = waste resource & inkonsisten.

📝 Tugas

Deploy First Model

Train sederhana scikit-learn model (mis. iris classifier).
Save model dengan joblib/pickle.
Build FastAPI endpoint /predict yang load model & return prediction.
Containerize dengan Docker.
Deploy ke cloud (Hugging Face Spaces, Railway, atau Render — semua punya free tier).
Test endpoint dengan curl atau Postman.
Bonus: tambahkan logging, basic auth, rate limiting.

Rangkuman

MLOps mengelola 95% dari ML system yang bukan kode model.
Pipeline standar: Collect → Validate → Engineer → Train → Eval → Deploy → Monitor.
Tooling: DVC, MLflow, Airflow, FastAPI, Evidently — pelajari minimum 2-3.
3 drift wajib monitor: data, concept, performance.
Deployment safety: shadow → canary → A/B → full rollout.
Tanpa platform MLOps, ML di scale = technical debt eksponensial.