On Mon, Mar 28, 2016 at 10:54 AM, Daniel J Walsh <dwa...@redhat.com> wrote:
> > > On 03/28/2016 10:45 AM, Eric Paris wrote: > >> On Mon, 2016-03-28 at 09:27 -0400, Colin Walters wrote: >> >>> Hi, >>> >>> In some of my use cases I have OpenShift/Kubernetes clusters that are >>> primarily certified on 1.9, and so I'd like to keep using that. But >>> it'd be useful to be able to quickly try out 1.10 on some of my >>> nodes, or in cases outside of a Kube cluster. >>> >>> Does anyone have any thoughts on this? What if we had two Docker >>> packages, and used a config option to determine which to run as >>> `docker.service` ? >>> >> I personally see little value in the engineering effort required to >> install two versions at once. There just don't seem to me to be many >> practical use cases. Actual users, people who need to get work done >> with docker will not want to constantly flip back and forth. Docker >> developers aren't like to want to do so as they are likely to just want >> to work on the latest and greatest. >> >> Which leaves me only able to think of the one oddball case you already >> pointed out. Where kubernetes only claims support for 1.9 even though >> Docker has released 1.10. Where a user may want to evaluate using plain >> docker vs using kubernetes. In this situation you either have the case >> where the user needs a feature in $KUBE_DOCKER_VERSION + 1, and they >> both cannot use kube and cannot use $KUBE_DOCKER_VERSION or they do not >> need a docker feature in $VERSION+1 in which case they should be able >> to do the comparison with only 1 version installed. >> >> So I don't see a real reason to need 2 versions installed from a user >> story point of view. And the fact that there is only one version of >> docker supported in a fedora release at a time leads bolsters my >> feelings like this is not something a user would need/want. In which >> case it makes me ask 'obviously I'm missing the point, why is someone >> asking for this?' >> >> I can only assume it is because of the pain involved in changing the >> version of a package when using rpmostree as a developer. Which makes >> me ask, 'should we be using atomic in this case?' When the explicit use >> case is about quickly iterating between two versions of packages and >> rpmostree is about entire images, it just seems like we have an >> impeedance mismatch which maybe shouldn't be 'solved'... >> >> -Eric >> >> Well a couple of points, we are currently blocked from pushing > docker-1.10 into fedora 23, because > it will break k8s, and it looks like we could have the same problem when > we go to ship docker-1.10 > into rhel in May. What issues arise with Docker 1.10 and Kubernetes? > If we shipped both in the same package a user with k8s could easily > switch to that version > and we could allow people in Fedora to use the latest version. > > On atomic host, we have been told that the two use cases of using it for > latest docker versus using it with > k8s and OpenShift need to be supported. Which is why we want to ship one > version to support both. > > It would be nice if docker would slow down enough for k8s to stay up, but > docker is already in rc2 for docker-1.11. > >
