Ceph

Yuri, Kraken was EOL in August 2017. why were you testing the upgrade path from kraken to luminous?

Still see in http://qa-proxy.ceph.com/teuthology/teuthology-2018-02-07_05:50:01-upgrade:client-upgrade-kraken-luminous-distro-basic-ovh/2164503/teuthology.log

The command causing this error is: 2018-01-03T07:57:45.906 INFO:teuthology.orchestra.run.ovh042:Running: "sudo yum -y install '' libcephfs_jni1 rbd-fuse ceph-radosgw librbd1 ceph-fuse python-ceph f ceph-mgr ceph librados2 cephfs-java libcephfs2 libcephfs-devel"

This is, essentially, the list of packages to be upgraded. If ceph-test is added to the list, the error will go away?

Ah, ceph-test is explicitly excluded from the list of packages:

  upgrade_workload:
    sequential:
    - install.upgrade:
        client.0: null
        exclude_packages:
        - ceph-test
        - ceph-test-dbg
        - libcephfs1

  sha1: 840cc7b1747fe12e42e7e182afd8298d1f8684b0 <-- luminous
...
  suite_sha1: ad30823c0d7544205669d09d1a6b0279f29ceab4 <-- kraken (!)

So this test is run with

--suite-branch kraken

@Nathan Weinberg

We do need to use `--suite-branch` as we agreed to be using clients from the appropriate named branches

https://github.com/ceph/ceph/tree/jewel/qa/suites/upgrade/client-upgrade-jewel/jewel-client-x
https://github.com/ceph/ceph/tree/kraken/qa/suites/upgrade/client-upgrade-kraken/kraken-client-x
https://github.com/ceph/ceph/tree/hammer/qa/suites/upgrade/client-upgrade-hammer/hammer-client-x

there also was a technical reason why we had to do so, but I can't recall ATM (IIRC we wanted to install correct packages on older clients w/o problems)

we still need to add: ceph-{base,osd,mon} >= 12.2.2.8 conflict with ceph-test < 12.2.2-8

because in kraken, ceph-test requires ceph-common without tying it with any specific version. so if we install, for instance, ceph-base (>= 12.2.2.8) to a system with ceph-test (kraken), we have no way to remove the conflicting old ceph-test.

but we are still facing a dilemma here: the test at https://github.com/ceph/ceph/blob/kraken/qa/suites/upgrade/client-upgrade-kraken/kraken-client-x/basic/1-install/kraken-client-x.yaml#L9 needs to preserve the old ceph-test for verifying that the kraken client still behaves when working with luminous.

to address this issue, probably we can add all the conflicting packages, namely ceph-{base,mon,osd}, to the exclude list of client.0 's install.upgrade task. because client.0 only hosts the old clients in this very test.

in short, two tasks:

• in master and luminous: ceph-{base,osd,mon} >= 12.2.2.8 conflict with ceph-test < 12.2.2-8
• in kraken: exclude ceph-{base,mon,osd} also.

@Yuri, ok we want to test {hammer,jewel,kraken} clients against luminous clusters but these tests don't actually do that:

https://github.com/ceph/ceph/tree/jewel/qa/suites/upgrade/client-upgrade-jewel/jewel-client-x
https://github.com/ceph/ceph/tree/kraken/qa/suites/upgrade/client-upgrade-kraken/kraken-client-x
https://github.com/ceph/ceph/tree/hammer/qa/suites/upgrade/client-upgrade-hammer/hammer-client-x

What they do is install a {hammer,jewel,kraken} cluster (including client) and then upgrade the client to luminous.

@Nathan Weinberg

After thinking about it and @jdurgin reminded that since we have workloads running before upgrading the client it seems we have all what we need. Agree?

@Yuri Since the test is fixable, we might as well keep running it :-)