Because of an unreliable hardware clock: collectd starts, writes a timestamp to the RRD files; timesyncd (the NTP client) sets the time back an hour; collectd's rrdtool plugin gets confused because time just went backward; collectd fails to log anything until the original logged timestamp is reached.

Before the most recent reboot, I had systemd order collectd.service After=time-sync.target; but that turned out not to actually work because it appears that timesyncd is allowing time-sync.target to be reached before it has actually synced the time.

I've opened a bug against timesyncd upstream: https://github.com/systemd/systemd/issues/5097