I tried this with real data (Satellite and Columbus module telemetry). At least for the timestamps, it is always a big win, because values are typically being s

I tried this with real data (Satellite and Columbus module telemetry). At least for the timestamps, it is always a big win, because values are typically being sampled at a precisely or at least somewhat regular frequency. For sample values, it is not as clear cut. The best way I found is to just try both approaches and use the one that results in better compression, if you can afford it. But for values I usually don't bother.