I probably shouldn't be surprised that collectd's memory cache takes quite a long time to drain to disk after turning off the aggregate 1Mbps of statistics streams which have been pouring into it for many years now. It must be coming up to half an hour now since I turned off the last incoming stream and disk usage still hasn't slacked off. Kind of impressive how well the caching works to consolidate updates to any given rrd file (which therefore must reduce SSD wear quite a lot), but it does mean restarting the collectd daemon managing a centralised repository can take a *long* time and cause quite a big gap in the records.