On 2020-09-09 15:51, Eugen Block wrote:
Hi Simon,
What about the idea of creating the cluster over two data centers? Would it be possible to modify the crush map, so one pool gets replicated over those two data centers and if one fails, the other one would still be functional?
A stretched cluster is a valid approach, but you have to consider several things like MON quorum (you'll need a third MON independent of the two DCs) and failure domains and resiliency. The crush map and rules can be easily adjusted to reflect two DCs.
There is a PR open to _explicitly_ support stretch clusters in Ceph [1]. To get it explained you can watch this presentation Gregory Farnum gave at FOSDEM 2020 [2]. Fortunately you can pause the video as he is going quite fast ;-). Better than two DCs are three DCs: that just works. And even better than theee are of course four DCs ... so you can recover from a complete DC failure ... Gr. Stefan [1]: https://github.com/ceph/ceph/pull/32336 [2]: https://archive.fosdem.org/2020/schedule/event/sds_ceph_stretch_clusters/