Proactive Cold Start Mitigation in Serverless Environments

A comprehensive machine learning pipeline for predicting and mitigating cold starts in serverless cloud computing environments using the Huawei Public Cloud Trace 2025 dataset.

📋 Overview

This project implements a sophisticated 31-day machine learning pipeline using PyTorch and scikit-learn to:

• Predict cold starts in serverless environments with high accuracy
• Forecast container cold start rates at multiple time horizons (1-min, 5-min, 15-min)
• Optimize resource allocation through an adaptive threshold controller
• Ensemble multiple models for improved prediction accuracy and robustness

Dataset: Huawei Public Cloud Trace 2025 – Region 1 Cold Start Traces

🏗️ Architecture

The system is built on a hybrid architecture with five key components:

Component	Technology	Purpose
Bidirectional LSTM	PyTorch	Captures sequential temporal patterns in cold start events
RandomForest Classifier	scikit-learn	Extracts insights from engineered tabular features
Confidence-Weighted Ensemble	Custom	Combines predictions weighted by validation AUC scores
Adaptive Threshold Controller	Custom	Dynamically adjusts decision thresholds based on FP/FN trade-offs
Multi-Horizon Predictor	PyTorch + scikit-learn	Generates forecasts at 1-minute, 5-minute, and 15-minute intervals

📊 Data Structure

The pipeline processes 31 days of continuous Huawei cloud trace data with explicit train/validation/test splits:

Dataset	Days	Records	Purpose
Training	0–18	19 days	Model training and parameter optimization
Validation	19–24	6 days	Hyperparameter tuning and model selection
Test	25–30	6 days	Final held-out evaluation and performance assessment

Per-Minute Records Include:

• total: Total function invocations
• cold: Cold start count
• cold_rate: Cold start ratio (cold / total)
• Temporal features: hour of day, day of week, business hours indicator
• Aggregation features: 5-min, 15-min, 60-min rolling means, standard deviations, and cold rates
• Lagged features: 1-min, 5-min, 15-min, 30-min lagged totals and cold counts
• Derived features: trend indicators, burst ratios, cold acceleration metrics

🚀 Installation

Prerequisites

• Python: 3.14 or higher
• Package Manager: pip or uv (recommended for speed)

Setup Steps

1. Clone the repository and install dependencies:

# Using uv (recommended)
uv sync

# Or using pip
pip install -r requirements.txt

2. (Optional) Install GPU support for CUDA 12.1:

# Using uv
uv pip install torch --index-url https://download.pytorch.org/whl/cu121

# Or using pip
pip install torch --index-url https://download.pytorch.org/whl/cu121

3. Download and extract the dataset:

The dataset consists of 31 CSV files (day_0.csv through day_30.csv) representing one month of Huawei cloud traces.

Download: R1 Dataset

Extract:

unzip R1.zip

This creates a directory with CSV files named day_0.csv, day_1.csv, ..., day_30.csv.

▶️ Usage

Running the Full Pipeline

# Using uv
uv run python main.py --dir /path/to/data/

# Using python
python main.py --dir /path/to/data/

Command-Line Arguments

Argument	Description	Example
--dir	Path to directory containing extracted CSV files	./R1/ or /data/traces/

Example

uv run python main.py --dir ./R1/

Pipeline Outputs

The pipeline generates comprehensive results in the following directories:

• models/ – Trained model artifacts (BiLSTM, RandomForest, ensemble weights)
• results/ – Predictions, evaluation metrics, and performance visualizations
• Console output – Training logs, validation metrics, and final performance reports

Key metrics reported:

• ROC-AUC scores for each horizon and ensemble
• F1, Precision, and Recall scores
• ROC curves and threshold analysis plots
• Prediction accuracy on held-out test set

⚙️ Configuration & Customization

Hyperparameters

Edit main.py to modify:

LSTM_EPOCHS = 120          # Number of training epochs
BATCH_SIZE = 64            # Batch size for LSTM training
PATIENCE = 20              # Early stopping patience
COLD_THRESHOLD = 0.25      # Cold start classification threshold
SEQUENCE_LEN = 30          # LSTM input sequence length (in minutes)
PRIMARY_HORIZON = 'target_5min'  # Primary prediction horizon

Feature Configuration

Available prediction horizons:

• target_1min – 1-minute ahead forecast
• target_5min – 5-minute ahead forecast (default)
• target_15min – 15-minute ahead forecast

Modify PRIMARY_HORIZON in main.py to switch horizons.

Feature Engineering Details

The pipeline automatically generates:

Feature Category	Examples
Temporal	hour_of_day, day_of_week, is_business_hours
Aggregation	rolling_mean_5m, rolling_std_15m, cold_rate_60m
Lagged	lag_1m_total, lag_5m_cold, lag_30m_total
Derived	trend, burst_ratio, cold_acceleration

📈 Performance Metrics

The pipeline evaluates models using standard ML metrics:

Metric	Definition	Use Case
ROC-AUC	Area under receiver operating characteristic curve	Overall model discrimination ability
F1 Score	Harmonic mean of precision and recall	Balanced performance metric
Precision	TP / (TP + FP)	Minimize false alarms
Recall	TP / (TP + FN)	Minimize missed cold starts

Results are generated for each prediction horizon individually and for the ensemble model.

🛠️ Technical Stack

Component	Library	Version	Purpose
Deep Learning	PyTorch	≥2.2.0	LSTM model and training
ML & Metrics	scikit-learn	≥1.3.0	RandomForest, evaluation metrics
Data Processing	pandas	≥2.0.0	DataFrames and feature engineering
Numerical Computing	NumPy	≥1.24.0	Array operations
Visualization	Matplotlib	≥3.7.0	Plots and reports
Gradient Boosting	XGBoost	≥1.7.0	Optional ensemble component
Gradient Boosting	LightGBM	≥4.0.0	Optional ensemble component
Deep Learning	TensorFlow/Keras	≥2.13.0	Optional alternative to PyTorch

💻 Device Support

The system automatically detects and uses the optimal compute device:

• CUDA GPU (if available) – Recommended for faster training and inference
• CPU (fallback) – Works on any machine, slower but functional

GPU usage is logged during training. Check logs for CUDA device messages.

📁 Project Structure

serverless_cold_start/
├── main.py                          # Main pipeline orchestrator
├── requirements.txt                 # Python dependencies
├── pyproject.toml                   # Project metadata and config
├── README.md                        # This file
├── uv.lock                          # Dependency lock file (for uv)
├── config/
│   ├── config.py                    # Configuration constants
│   └── __init__.py
├── data_loader/
│   ├── loader.py                    # Dataset loading and parsing
│   └── __init__.py
├── feature_engineer/
│   ├── engineer.py                  # Feature extraction and engineering
│   └── __init__.py
├── models/
│   ├── bilstm_model.py             # BiLSTM PyTorch model definition
│   ├── adaptive_threshold_controller.py  # Threshold adaptation logic
│   ├── simulator.py                 # Simulation and evaluation
│   └── __init__.py
├── visualization/
│   ├── plots.py                     # Plotting and visualization functions
│   └── __init__.py
├── generated_models/                # Output directory for trained models
├── results/                         # Output directory for evaluation results
├── R1.zip                           # Dataset archive (after download)
└── R1/                              # Extracted dataset (day_0.csv through day_30.csv)

🔧 Troubleshooting

"day_0.csv not found" Error

Solution:

• Ensure the R1.zip dataset is downloaded and extracted: unzip R1.zip
• Verify the --dir path points to the correct extracted directory containing CSV files
• Confirm files are named exactly day_0.csv, day_1.csv, ..., day_30.csv

Download Fails

Solution:

• Check your internet connection
• Verify the Google Drive link is accessible and not quota-limited
• Try alternative download methods (browser, curl, wget)
• Check available disk space (requires ~500MB for dataset)

CUDA Out of Memory Error

Solution:

• Reduce BATCH_SIZE in main.py (try 32 or 16)
• Switch to CPU by using non-CUDA PyTorch: pip install torch
• Use a GPU with higher VRAM or split the data into smaller chunks

Slow Training / Low GPU Utilization

Solution:

• Verify CUDA is properly installed: python -c "import torch; print(torch.cuda.is_available())"
• Check training logs for CUDA device usage
• Increase BATCH_SIZE if GPU memory allows (up to 256 or 512)
• Use a machine with better GPU hardware (A100, H100, RTX 4090)

Import Errors (TensorFlow, XGBoost, etc.)

Solution:

• Reinstall all dependencies: pip install -r requirements.txt --force-reinstall
• Verify Python version is 3.14+: python --version
• Check for version conflicts in pyproject.toml (TensorFlow compatibility constraints)

📚 References

• PyTorch Documentation: https://pytorch.org/docs/stable/
• scikit-learn Guide: https://scikit-learn.org/stable/
• Huawei Cloud Traces: https://github.com/sir-lab/data-release/blob/main/README_data_release_2025.md
• LSTM/BiLSTM Theory: https://colah.github.io/posts/2015-08-Understanding-LSTMs/
• Ensemble Methods: https://scikit-learn.org/stable/modules/ensemble.html

❓ Support & Questions

For issues, questions, or feature requests:

1. Check the Troubleshooting section above
2. Review existing GitHub Issues
3. Create a new issue with:
   • Error message and full traceback
   • Python version and OS
   • Steps to reproduce the problem
   • Hardware specs (GPU/CPU, RAM)