3 ms·
I'm curious what the distribution of time offsets between servers running NTP is. Properly functioning NTP should maintain an offset well under 100ms, but what
by aidenn0 7y ago
I'm curious what the distribution of time offsets between servers running NTP is. Properly functioning NTP should maintain an offset well under 100ms, but what fraction of the time servers are "properly functioning" is a question I'm interested in knowing.
- thedance 7y agoThis is really the key point, and the main thing that many people overlook. The purpose of TrueTime in Spanner is not gained by absolute accuracy. It's gained by knowing the clock uncertainty across all participants and affirmatively ejecting any participant with excessive uncertainty. The atomic clocks are totally besides the point.
- hinkley 7y agoBut the atomic clock is how they keep the throughput from crashing every time a corner case is encountered.
- hermitdev 7y agoPTP (Precision Time Protocol) can get you sub microsecond synchronization on a LAN. See: https://en.wikipedia.org/wiki/Precision_Time_Protocol https://en.wikipedia.org/wiki/Precision_Time_Protocol Most recent spec is IEEE 1588-2019.
- aidenn0 7y agoNTP is more than good enough to get 7ms synchronization on a LAN when working correctly, so the fact that PTP can do even better is perhaps uninteresting in light of the fact that NTP often fails to get anywhere near its theoretical performance (other comments mention clocks off by years).
- sethammons 7y agoTalking with ops folks at work, they have seen system clocks in our fleet (which run ntpd) suddenly report times off by _years_ and then go back to "accurate." We have thousands of nodes for the software I work on. We sometimes get negative times when comparing events, if only by a few seconds. You can't fully trust time in a distributed system.
- aidenn0 7y agoSure, but it's also possible to write a word out to memory and then read something different back. There are probably many DBs that fail to meet their requirements when that happens. My sense is that time errors are much more common, but I'm interested in actual data surrounding that.