Problem
Profiling a prediction request shows something surprising: the model itself barely registers, but the online feature lookup; fetching the dozens of precomputed feature values the model actually needs; dominates the response time. It's been treated as an afterthought, sized for whatever fell out of the default config. Size the online feature cache for the real read/write volume it's actually under, not the volume nobody bothered to check.
Required components
Click to add a component to the canvas.
Additional components
Browse by category. Click or drag a component onto the canvas.