MTFSA-Autoscaler

Meta + Few-Shot + Transfer Learning Hybrid Autoscaler for FaaS Cold Start Prediction

Predicts function-level cold starts using only 3 hours of historical data — no fine-tuning required.
Achieved RMSE 0.117 | R² 0.876 | 100% accuracy within ±0.5 on 1,272 sparse functions across 5 regions.

Live Dashboard → https://mtfsa-autoscaler.onrender.com

---

Key Features

• 3-shot prediction with Meta-Few-Shot LSTM pre-trained on dense region (R1)
• Random Forest refinement layer for sub-cold-start precision
• Real-time interactive dashboard with scaling alerts
• Tiny models: 77 KB .keras + 41 KB .pkl
• Fully cloud-native deployment on Render (free tier)

Results (31-day Huawei Cloud traces)

Metric	Value
RMSE	0.117
R² Score	0.876
Accuracy (±0.5 cold starts)	100%
Prediction Horizon	Next hour
Training Data Needed	~3 hours

Architecture

                 ┌─────────────────────┐
                 │  Dense Region (R1)   │
                 │   377 functions      │
                 └─────────▲───────────┘
                           │
               Meta-Few-Shot Pre-training (LSTM)
                           │
                           ▼
            Transfer to Sparse 1,272 Sparse Functions (R2–R5)
                           │
                           ▼
                    3-shot Inference
                           │
                           ▼
                     LSTM Prediction
                           │
                           ▼
                  Random Forest Refiner
                           │
                           ▼
               Final Forecast + Auto-Scaling Alert

Project Structure

├── app_final.py
├── models/
│   ├── final_model2.keras    (77 KB)
│   └── rf_refiner.pkl        (41 KB)
├── data/processed/           (Parquet traces)
├── requirements.txt
└── render.yaml               (One-click Render deploy)

Quick Start

• git clone https://github.com/Saanvim11/MTFSA-Autoscaler.git
• cd MTFSA-Autoscaler
• pip install -r requirements.txt
• Render run app_final.py