> I guess it occurs because ntpd does not poll constraints servers as often > as ntp peers.
I've sees this proposed before. Why would it need to poll again? The constraints servers gave it a (latency shifted, but less than 1s since internet doesn't reach mars yet), which we can compare against the time of the machine pre-adjusted. Now we have a baseline that will never be untrue. ntpd can now shift the baseline, to make it more accurate. It ends up getting CLOSER to the constraint time, not further. If it is getting further, your OS/machine interaction is broken. That's the bug to fix. constraints should not be blamed for this.
