saurin1004 / aws-face-recognition-autoscaler

Cloud-based face recognition system using Python/Flask that auto-scales EC2 instances (0-15) via custom logic. Uses SQS for decoupled messaging, S3 for storage, and a PyTorch model. Handles 100+ requests at <1.2s latency while cost-optimizing through scale-to-zero architecture. Built with Boto3 without AWS Auto Scaling service.

Auto-Scaling Face Recognition System on AWS

A cloud-native application that dynamically scales compute resources for face recognition workloads using AWS services (EC2, S3, SQS) with custom auto-scaling logic.

Key Features

• Elastic Infrastructure: Automatically scales from 0 to 15 EC2 instances based on SQS queue depth
• Cost-Efficient: Scales to zero when idle using custom logic (no AWS Auto Scaling)
• High Throughput: Processes 100+ concurrent requests with <1.2s average latency
• Fault-Tolerant: Decoupled architecture using SQS queues between web/app tiers

Architecture

graph TD
    A[Client] -->|Upload Image| B[Web Tier]
    B -->|Store| C[(S3 Input Bucket)]
    B -->|Request| D[SQS Request Queue]
    D -->|Connection| E[App Tier]
    E -->|Controller auto scales| E
    E -->|Process| F[Face Recognition Model]
    E -->|Store Result| G[(S3 Output Bucket)]
    E -->|Response| H[SQS Response Queue]
    H -->|Connection| B
    B -->|Return Result| A

Loading

Tech Stack

• Compute: EC2 (Web Tier + Custom Auto-Scaling App Tier)

• Storage: S3 (Input/Output Buckets)

• Messaging: SQS (Request/Response Queues)

• Languages: Python (Flask, Boto3)

• Machine Learning: Pretrained Face Recognition Model (PyTorch)

Setup Instructions

1. Prerequisites

pip install boto3 flask torch torchvision

2. Deploy Infrastructure

• S3 Buckets:

  • Create input bucket: in-bucket

  • Create output bucket: out-bucket

• SQS Queues:

  • Create request queue: req-queue

  • Create response queue: resp-queue

• EC2 Configuration:

  • Launch instances with IAM roles granting access to S3 and SQS

  • Configure security groups to allow necessary traffic

3. Launch Services

# Web Tier (Flask server)
nohup python3 server.py &

# Autoscaling Controller
nohup python3 controller.py &

NOTE: For App-tier, I have configured EC2 instance using Systemd command so that the backend.py gets initiated as soon as an app-tier instance is set to Running

Performance Metrics

Performance Metrics

Metric	Value	Description
Max Throughput	100 req/sec	Sustained request rate under load testing
Avg Latency	0.96s	End-to-end processing time (image upload → recognition result)
Scale-Out Time	<5 sec	Time to launch new App Tier instances when queue depth increases
Scale-In Time	<2 sec	Time to terminate idle instances when queue is empty
Error Rate	<0.1%	Failed requests during stress testing

Testing Methodology

• Tool: Custom python script
• Workload: 1000 face images (512x512px)
• Environment: AWS us-east-1 (t2.micro instances)