A hybrid approach on aws lambda for cost optimization in serverless architecture

In serverless architectures, various optimization techniques have been developed that accelerate machine learning inference and reduce operational costs. Within this context, the batching approach has emerged as a particularly effective strategy. Experimental studies conducted on computer vision models such as ResNet-v2, ResNet-18, ResNet-50, Inception-v4, and MoBiNet report that dynamically adjusting batch sizes and timeout thresholds can reduce costs by approximately 50\% while significantly decreasing queueing delays. Conversely, studies leveraging predictive resource management and vertical/horizontal scaling techniques have achieved performance improvements in the range of 50–95\% and up to 95\% Service Level Objective (SLO) compliance on models such as ResNet-50, InceptionV3, Xception, and Wavenet. Collectively, these findings underscore the critical importance of both batching and resource scaling for enhancing performance and efficiency. This study proposes an innovative hybrid method that integrates batch processing with concurrency. The methodology involves a controlled comparison on AWS Lambda between a non-optimized baseline scenario (50 concurrent threads processing 5,000 individual requests) and the hybrid approach introduced in this thesis (a batch size of 5 with 350 concurrent threads processing 1,000 requests). Key performance metrics, including total execution time, latency distribution, and system efficiency, were systematically assessed. Experimental results demonstrate that the proposed hybrid approach reduced total execution time from 139.97 seconds to 30.30 seconds, achieving a substantial performance improvement of 78.35\% over the baseline. Furthermore, the approach was shown to mitigate queuing delays and enhance overall system efficiency. By offering a practical and quantitatively validated solution for cost-effective and high-performance execution of machine learning inference tasks in serverless environments, this study makes a significant contribution to the literature.