Google wrote a blog post about this a few years ago:

- http://googleblog.blogspot.com/2011/09/time-technology-and-l...

There was a post in the past that google have their own custom timeservers for their servers, slowly adapting time over multiple days and months, so they are not affected by this.

In their databases that depend on time, they repeat that they use an external source of time.

It seems like some of their designs are constrained around always non negative deltas between events, making things append only so their trees or hash tables are only balanced once when written to disk.

I haven't read about Spanner, but wouldn't it be possible for their clock to measure time based on an epoch, instead of using the system based on physical factors (the rotation of the earth around its axis and around the sun) that we use?

It is probably easier in the long run if the servers had the same time as the machines that people use, which are regularly updated by their OS's choice of NTP server, since people run reports and react accordingly.

It wouldn't affect Spanner or any other sanely implemented system - usually nowadays time is saved as "units from epoch", and leap seconds, days, etc. don't matter for internal calculations. It matters only for final conversion to printable format, debug logs, etc.

Most UNIX systems store time as "seconds from the epoch, according to UTC". The advantage is that you can always assume that 1 day = 86400 seconds. The disadvantage is that the time is affected by leap seconds. The same-numbered POSIX second can be repeated twice in a row.

Spanner instead deals with it by changing the length of the second.

If you're assuming time intervals are accurate to the second, or that times are monotonic, have you actually reconfigured your system to use TAI?

One cannot even assume that '1 day = 86400 seconds' because of DST.

I've never seen a UNIX system that changes the system clock due to DST. You change the local time zone.