14 May
2020
14 May
'20
8:35 a.m.
Den ons 13 maj 2020 kl 22:37 skrev Bryan Henderson <bryanh(a)giraffe-data.com
:
I'm surprised I couldn't find this explained
anywhere (I did look), but ...
What is the pgmap and why does it get updated every few seconds on a tiny
cluster that's mostly idle?
I was sure it was updated exactly once per second.
I do know what a placement group (PG) is and that when
documentation talks
about placement group maps, it is talking about something else -- mapping
of
PGs to OSDs by CRUSH and OSD maps.
I thought it was a method (the method?) to know if a PG comes back from a
crashed OSD/host, to know if it was up-to-date or old since it would have
an older timestamp.
Since using computer time and date is fraught with peril, having the whole
cluster just bump that single number every second (and writing it to the PG
on each write) would allow a mostly idle PG that comes back after an hour
of unexpected downtime to easily know if it needs no recovery, a little bit
of delta to get up-to-date or a full copy from the primary in order to
become a part of the replica set for that PG.
--
May the most significant bit of your life be positive.