It seems like it probably isn't necessary in this example to process an image every 15 seconds. You could simply scan every other image or one image per minute maybe. I wonder what portion of time-constrained problems could actually be helped by a similar techniques of sampling.
From lecture, the latency of image processing is actually important for this use case. Some astronomical events may happen very quickly so its important for the LSST to quickly process an image and send that data to telescopes around the world (there is latency here too!). With a minute of latency, the entire event may be missed!