1. Introduction
Real-time video analytics workloads such as retail footfall counting are often bursty by nature, tied to business hours or events, making the pay-per-invocation model of serverless computing attractive despite known concerns about cold-start latency for compute-heavy, GPU-dependent inference tasks.
2. Methodology
An identical YOLOv8-based object detection video analytics pipeline was deployed on AWS Lambda with container image support, Google Cloud Run, and a baseline Kubernetes deployment on reserved GPU nodes, and benchmarked under three synthetic traffic patterns, constant load, business-hours load, and bursty event-triggered load, measuring end-to-end frame processing latency, monthly infrastructure cost, and cold-start frequency.
3. Results
The Kubernetes baseline achieved the lowest steady-state per-frame latency at 82 milliseconds versus 134 milliseconds for Cloud Run and 151 milliseconds for Lambda, but under the bursty event-triggered traffic pattern, Cloud Run and Lambda reduced monthly cost by 58 percent and 51 percent respectively relative to the always-on Kubernetes baseline, at the cost of an average cold-start penalty of 1.9 seconds affecting the first request after an idle period.
4. Conclusion
The choice between serverless and traditional deployment for real-time video analytics should be driven by traffic pattern, serverless platforms offer substantial cost advantages for bursty workloads that can tolerate occasional cold-start latency, while steady high-throughput workloads remain better served by reserved infrastructure. Future work will evaluate provisioned-concurrency options to mitigate cold-start penalties.
References
[1] Jonas E. et al., Cloud programming simplified: A Berkeley view on serverless computing, arXiv, 2019. [2] Jocher G. et al., YOLOv8, Ultralytics, 2023.