Re: I miss the dayxs when I could flip between 3 programs simultaneosuly.
I wonder how frequently they send those re-synch frames in the data stream?
This was talked about briefly in the article, the stream is arranged into GOP, Groups Of Pictures, a list of frames. Each frame can be of I, P, B. I-frames are the entire picture, P-frames are forward predictive and B-frames are bi-directionally predictive frames, ie they add to the previous/next frame. If you haven't decoded the previous frame, then it can only show the difference, and the video looks "corrupted".
Obviously, I-frames take the most space, P-frames, being one way predictive, take up less space, and B-frames take even less space, so for optimum quality at a given bandwidth, you want as few as possible I-frames, less P-frames and more B-frames. However, as you point out, you need as many I-frames as possible to make seeking/scrubbing not appear corrupted.
A typical GOP may have something like this structure "IBBPBBPBBPBBI". The next I-frame is 12 frames after the first I-frame, and this is the GOP length (N). The maximum distance between 2 reference frames (like an I or P frame) in the GOP is 3, which is the GOP size (M), so this GOP would be described as N=12 M=3. The GOP length is also called the Intra Period.
Anyway, all of these things are configurable by the encoder. So the answer to your question is, "as infrequently as the encoder thought they could get away with". You want enough I-frames that you can seek easily, you want as many P/B-frames as possible to keep the bitrate down. Pay your dollar, make your choice.