customize-failure-domain has no effect when ceph-mon is deployed in a container

[Nobuto Murata]

Let's say I have 3 machines among 3 racks.

node01 - rack01
node02 - rack02
node03 - rack03

And MAAS zones are configured as zone-rack01, 02, 03. In this case, it's quite common to deploy ceph-mon as containers on top of physical nodes and ceph-osd to the hosts. JUJU_AVAILABILITY_ZONE is correctly set as rack-01, 02, 03 for the hosts, but it's empty in containers. Therefore, customize-failure-domain block never be executed:
https://github.com/openstack/charm-ceph-mon/blob/master/hooks/ceph_hooks.py#L283-L308

As a result, no "step chooseleaf firstn 0 type rack" line in the crush map.

[container]
$ juju run --unit ceph-mon/0 'env | grep JUJU_AVAILABILITY_ZONE'
JUJU_AVAILABILITY_ZONE=

[physical host]
$ juju run --unit ceph-osd/0 'env | grep JUJU_AVAILABILITY_ZONE'
JUJU_AVAILABILITY_ZONE=zone-rack01

[juju debug-log -i unit-ceph-mon-2 --replay]
unit-ceph-mon-2: 15:33:28 INFO unit.ceph-mon/2.juju-log mon:1: Your Juju environment doesn'thave support for Availability Zones

[Chris MacNaughton (chris.macnaughton)]

This is a new version of a recurring Juju bug where the JUJU_AVAILABILITY_ZONE is not set on a specific machine type, it happened in general on all machines, and then again in MAAS when we were developing this feature.

[Tim Penhey (thumper)]

It looks as if the host machine availability zone is not propogated to the containers running on that host. This *should* be a simplish fix.

[Dmitrii Shcherbakov (dmitriis)]

http://paste.ubuntu.com/25411702/ - 2.2.x
http://paste.ubuntu.com/25411705/ - devel (2.3-alpha1)

PRs:
https://github.com/juju/juju/pull/7795
https://github.com/juju/juju/pull/7796

[Ante Karamatić (ivoks)]

For us this is somewhat urgent/critical...

[Dmitrii Shcherbakov (dmitriis)]

merged for 2.3: https://github.com/juju/juju/pull/7795

hopefully will get merged for 2.2.3 https://github.com/juju/juju/pull/7796

[Nobuto Murata (nobuto)]

Fixed in Juju stable release. Marking the charm target as Invalid.