On Wed, May 29, 2019 at 3:35 PM Patrick Donnelly <pdonnell@xxxxxxxxxx> wrote: > So we need an upgrade test in qa/suites/fs/upgrade/ that does this > testing. That will mostly be migrating the existing test in > fs/basic_functional/tasks/volume-client to an upgrade test. The tricky > part is installing luminous on the client machine and then configuring > it to talk to the Nautilus/Octopus cluster. Up to now, we normally do > testing of old clients by first setting up the cluster with the older > version we like to test and then upgrading every node except the > clients. This won't work anymore (easily) because we no longer can > upgrade directly from Luminous to master/Octopus. So, we need to get a > little smarter by simply installing luminous packages on a client node > and then installing a ceph.conf to talk to the Nautilus/Octopus > cluster. +Josh suggested that we could split out the function setting > up the ceph.conf in qa/tasks/ceph.py into a separate task that can be > run on the clients. It does need to be made a little smarter so that > it installs a ceph.conf that is readable by a luminous client. In > particular, it can't dump v2 monitor addresses to the ceph.conf. Adding on to this: Sage thinks we can just do the install luminous task on all nodes and then install.upgrade on everything but one of the clients. Finally, run the ceph task on all the nodes but add a few options to only use the v1 addresses so that the ceph task doesn't put a ceph.conf with v2 addresses on the client node. See also rados/thrash-old-clients. -- Patrick Donnelly, Ph.D. He / Him / His Senior Software Engineer Red Hat Sunnyvale, CA GPG: 19F28A586F808C2402351B93C3301A3E258DD79D
