[Mosu]

Yeah, even us Matroska developers agree that the choice of how to timestamp frames wasn't a good one, but one that will most likely never change as too much stuff out there wouldn't be able to handle anything else.

In order to put things into perspective: each cluster element contains at least the cluster timestamp. As a generalization you could say that at µs resolution you'd end up with roughly twelve bytes of overhead per 32ms of content. For a two hour movie you'd end up with ~ 2h * 60m/h * 60s/m * 1000ms/s * 12bytes/cluster / 32ms/cluster / 1024bytes/KB / 1024KB/MB = 2.5MB of additional overhead. Due to B frames and them being out of order, the actual overhead will often be higher (I just ran a test with a 2h movie where the additional overhead was more like 4.2MB while total file size was 1.6GB). Personally I wouldn't call either of those numbers a "huge overhead". My days of trying to fit movies on single CDs are long gone, luckily.