Title with apologies to Tom Robbins, because what I’m going to describe is more of a bad smell than a perfume, through initially an elusive one.
In NTP, samples from each local clock or peer (upstream NTP host) are accumulated into an array of length 8 - called, for some historical reason I don’t know, the peer’s "shift register" . The old samples are kept around so ntpd can do statistical smoothing on the time series, tossing outliers, before (possibly) using the latest one to update the clock.
The quality of a peer is measured by its jitter. This is recomputed each time the sample-processing code is called as the RMS (root mean square) of distances from the median sample. Think of it as a running estimate of the amount of noise in the time series.
The code sees peers with low jitter as low-noise - better - and weights them higher in the arcane formulae that derive consensus UTC time from all the incoming samples.
Gary Miller, who is far better at reading an ntpq display than I am or probably will ever be, noticed that something in the code seemed occasionally to be zeroing peer jitter values. Telling us, in effect, that the affected peer is noiseless and should be weighted highly.
I didn’t believe this at first. Jitter gets zeroed in exactly three places, and we can count one of them out because it’s the local-clock driver. If you’re going to swallow your tail by using your system clock as a peer source (which you might want to do in some unusual deployment cases), you’d best assume it’s accurate.
In the two other places, it looked like jitter was zeroed only just before it is recomputed from the contents of the shift register. Only…it turns out that if the shift register from a network peer is empty (all samples have been filtered out) it was possible for that recomputation not to take place.
So, we had a worst case: bad samples come in, jitter is zeroed, samples get all tossed out, jitter is not recomputed - and the clock with the bad samples looks like the best possible source when the next sample comes in.
The section of code this bug lived in - the peer-sample processor in the central protocol machine - is, I believe, very old and very stable. It seems likely that the bug had been lurking there for more than twenty years.
I’m not sure why it was never spotted before, but it might help account for a general flakiness in ntpd’s behavior near startup that has been often observed but never properly characterized or explained. Close to startup (less than 8 sampling intervals from it) the shift register hasn’t yet seen enough samples for good ones to fill it yet, and it might have been easier to trigger the zero-sample error case.
It will take some time for the results of this fix to emerge from the fog of uncertainty around NTP performance. I think the strongest single possibility is that it will speed up convergence after startup in conditions of bad network weather.
It’d be nice to know for sure, but reasoning about NTP’s behavior is hard. The core algorithms are both recondite and surrounded by a lot of complicated moving parts.
Almost the best you can do with a hairball like the NTP reference code is (1) get good at spotting bugs like this that break logical invariants ("jitter is always recomputed to a nonzero value after being zeroed") and fixing them. And (2) gradually comprehend its intended behavior through observation and experiment. In some ways it’s more like exploring an unknown continent than writing conventional code.