A retail chain wants each checkout camera to flag an unscanned item before the shopper steps away. The computer on that camera has little memory, a tight power budget, and only a brief moment to answer. Which constraint should decide how this model is deployed?

The camera computer has to decide before the shopper leaves, using only the memory and power that sit on that device. Deployment at the edge is a fit problem: the model that runs there must stay inside those three limits, even if a larger model was trained somewhere else. Parking the full training cluster beside every lane treats a training farm as a checkout device and ignores heat, power, and space at the register. A distant nightly batch can run a large model, but the flag arrives after the shopper is gone, so it misses the latency the lane actually has. Paging an unchanged oversized model from a slow disk does not create memory the camera computer lacks, and the extra reads make the brief decision window harder to hit. Exam caveat: an edge constraint governs the machine that must answer in place; it does not forbid training the model earlier on a larger system. Operational check: load the candidate on a checkout camera computer and record memory use, power draw, and the time from frame to flag, and ship it only when all three sit inside that device's limits.