Why Amazon Changed the Architecture of Prime Video's VQA Service.

// Old Architecture: Intermediate storage hops via S3
// Each detector ran in a separate AWS Lambda function

public class VideoDetectorLambda {
    private final AmazonS3 s3Client;

    public void handleRequest(S3Event event) {
        // Step 1: Download video frames from S3 (High Network I/O & latency)
        byte[] videoFrame = s3Client.getObject("vqa-frames-bucket", "stream-101/frame-45.jpg");

        // Step 2: Run quality analysis algorithm
        DefectResult result = analyzeDefect(videoFrame);

        // Step 3: Write intermediate results back to S3 for next Step Function stage
        s3Client.putObject("vqa-frames-bucket", "stream-101/defect-result.json", result.toJson());
    }
}

// New Architecture: In-Memory Processing in a Single Process
// All detectors run together inside one container on Amazon ECS

public class VideoQualityAnalysisService {
    private final List<VideoDefectDetector> detectors;

    public void processStreamChunk(VideoChunk chunk) {
        // Frames stay in shared RAM memory (No S3 calls, zero network hops)
        byte[] frameBuffer = chunk.getRawBuffer();

        // Detectors process the buffer directly in-memory
        for (VideoDefectDetector detector : detectors) {
            detector.analyze(frameBuffer);
        }

        // Only final aggregated defect alerts are sent to storage/alerts
    }
}