Forecasting Knative Workload Traffic with Machine Learning to Minimize Cold Starts

Anil, Abigail Mariam (2025) Forecasting Knative Workload Traffic with Machine Learning to Minimize Cold Starts. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

On serverless platforms like Knative, scaling is automatically configured, however, the default reactive autoscaler reacts only after the load increases, leading to cold starts and latency spikes with bursty and irregular traffic that may not allow most cloud applications to operate efficiently. Previous research has investigated predictive autoscaling systems and hybrid forecasting of cloud systems, yet the end-to-end analysis of machine-learning-based forecasting as part of the Knative autoscaling loop using actual serverless traces has not been performed. This work builds a predictive autoscaling mechanism in Knative based on three predictive models, namely Prophet, LSTM and, Hybrid (Prophet-residuals-plus-LSTM) models, which were trained using the Azure Functions invocation data. Pre-scaling of pods is done based on the predictions before the demand arises. The experimentation results revealed that all the predictive strategies decrease the cold start latency in burst and transition phases. The experiments also show that the forecasting error for LSTM and Hybrid models is comparatively lower than that of Prophet for irregular traces. Prophet demonstrates strong performance for smooth and seasonal workloads, theoretically positioning ML forecasting as a safe extension, instead of being a replacement for reactive auto scaling in Knative. In practice, they maintain low tail latency for latency sensitive functions with relatively low additional resource cost. Overall, the evaluation shows that machine learning based predictive autoscaler reduces cold start latency in burst-heavy workloads without compromising its behaviour and is comparable to the Knative Pod Autoscaler with steady traffic, making it a complementary solution to reactive scaling.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Kazmi, Aqeel UNSPECIFIED
Subjects:	T Technology > T Technology (General) > Information Technology > Cloud computing Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Cloud Computing
Depositing User:	Ciara O'Brien
Date Deposited:	31 Aug 2026 12:20
Last Modified:	31 Aug 2026 12:20
URI:	https://norma.ncirl.ie/id/eprint/9690

Actions (login required)

View Item