150% throughput. 6M images an hour. In production.
The challenge
The provider's electronic shelf labels are only as good as the service that renders them, and that image-generation microservice was the bottleneck: expected to produce 1,200 images per second, it managed about 650. Under load it leaked events, distributed jobs unevenly, and burned resources it did not need — inconsistent output and a cost line growing faster than the customer base. For a platform billing on scale, every new retail customer added load the service could not absorb without degrading or over-provisioning. The scaling model did not match the bursty, event-driven nature of the workload.
Our approach
Ariviti redesigned the service around event-driven autoscaling with KEDA (Kubernetes Event-Driven Autoscaler), scaling on the actual event backlog rather than coarse infrastructure signals. The hard part was scale-down: KEDA does not handle in-application shutdown cleanly, so Ariviti implemented a Kubernetes preStop hook to drain in-flight work gracefully before a pod terminated, eliminating event loss. The result tracked its workload precisely — full capacity under burst, minimal footprint when idle, no events lost in either direction. Delivered through Ariviti's Virtual Technology Office (VTO) model: outcome-owned performance engineering against throughput and cost targets, proven in production.
The result
150% throughput improvement — from 650 to 1,600 images per second, 6 million images an hour. Significant infrastructure cost reduction from demand-matched capacity, and many more retail customers served per instance — in production.
