Exploring the Economics of Serverless: An Analysis of Cost Discrepancy in Auto-Scaling

Sammanfattning

In the realm of cloud computing, serverless computing emerges as a contemporary paradigm, distinguished by its utilization of auto- scaling—a dynamic resource allocation mechanism that operates seamlessly without manual intervention. Serverless architecture extends cloud computing by unburdening its user from resource management and allocation. It provides pay-as-you-go billing model, where a user only pays for the resources they consume. Cost efficiency is a critical factor in almost every successful serverless use case, and auto-scaling is meant to enable this by managing resources, avoiding overprovisioning, and scaling up or down based on the workload. Although serverless pricing appears straightforward, with a fixed cost per function invocation and a cost per time unit of function execution, it encompasses hidden expenses, notably stemming from factors like cold-start delays. These hidden costs may result in discrepancies between estimated and actual costs for users. Previous research has highlighted that the effective cost hinges on the arrival pattern of the serverless workload, implying the existence of hidden expenses linked to auto-scaling serverless deployments. This thesis investigates the hidden costs associated to auto-scaling through experiments conducted across various workload patterns on the Google Cloud Platform.