Hybrid processing splits the analytics work between the edge (camera or local device) and the cloud (or a central server), so each does what it is best at. The usual pattern is a triage cascade: a light, fast model on the edge watches everything and discards the boring majority, forwarding only the small fraction worth a closer look to the heavier model in the cloud. It is the architecture most real systems converge on, because the two extremes each waste something.
The division of labour is the point. Pure edge is fast and cheap but capped by on-device compute; pure cloud is powerful but expensive and slow to reach. Hybrid keeps the instant, high-volume filtering local and reserves the expensive reasoning for the events that earn it — commonly the edge passes only around 5% of frames upward, so the cloud sees a trickle, not a torrent. A store-and-forward seam (a local ring buffer plus buffered upload) lets the system keep recording and catch up after a network outage instead of losing footage.
The pitfall is drawing the boundary badly. Put too much on the edge and you miss what small models cannot see; lean too hard on the cloud and you pay the bandwidth and latency you were trying to avoid. The split should follow the job's latency tolerance and the cost of a miss, and it must handle the seam — what happens to in-flight data when the link drops — deliberately, not as an afterthought.
