Erik Lindahl <erik.lindahl@xxxxxxxxx> writes: >> On 20 Aug 2021, at 10:39, Nico Schottelius <nico.schottelius@xxxxxxxxxxx> wrote: >> >> I believe mid term everyone will need to provide their own image >> registries, as the approach of "everything is at dockerhub|quay" >> does not scale well. > > Yeah, this particular issue is not hard to fix technically (and I just > checked and realized there are also _client_ pull limits that apply > even to OSS repo). Yes, that is the problem we are running into from time to time. > However, I also think it's wise to take a step back and consider > whether it's just a matter of a technical mishap (docker suddenly > introducing limits) or two, or if these are signs the orchestration > isn't as simple as we had hoped. My personal opinion is still that focussing on something like rook for containers and leaving the base native would make most sense. If you are going down containers *anyway*, k8s is helpful for large scale deployments. If you are not, then completely containerless might be easier to document, teach and maintain. > If we need to set up our own container registry, our Ceph > orchestration is gradually getting more complicated than our entire > salt setup for ~200 nodes, which to me is an indication of something > not working as intended :-) Sorry, that's *not* what I meant: I think that each Open Source project (like Ceph) might need to setup their own registry like it happened with package repositories many years ago. Now, I am aware that the ceph team is working at redhat, redhat is driving quay.io, so the logically choice would be quay.io. But the problem with that is rate limiting on quay.io + the lack of IPv6 in our case, which makes all IPv6 hosts look like coming from one IPv4 address, which gives horrible rate limits. Even in the private IPv4 case that will be the same problem - dozens or hundreds of nodes are pulling from quay.io and your cluster gets rate limited. In practice that means that downloading a single ceph image might take hours, we have experienced this quite some times already. Thus an upgrade with cephadm or rook will potentially be delayed by hours, just for pulling in the image. Now you can argue that the users/consumers should carry some of the weight of providing the images and we from ungleich would be happy to sponsor a public image cache, if necessary. In short: the container registry move is not the only problem, the registry limits are a big problem for ceph clusters and require additional local caching at the moment. I would certainly prefer this being solved somewhere more nearby upstream instead of everyone running their on nexus/harbor/docker registry. Greetings from containerland, Nico -- Sustainable and modern Infrastructures by ungleich.ch _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx