contiv / build Goto Github PK

The task to update copy systemd units for docker(enable cluster store) in role docker introduces a dependency between the two roles.

We should conditionally run this task only when etcd role is going to be run to make this dependency more explicit.

right now the way ceph configuration is generated (as shown in the snippet from ansible/roles/ceph-common/templates/ceph.conf.j2 below) results in a dependency that all mons and osds hosts need to be configured together in single playbook run.

In a real cluster, we would need to allow incremental provisioning of new mons and osds. This issue tracks this requirement.

 79 {% for host in groups[mon_group_name] %}
 80   {% if hostvars[host]['ansible_hostname'] is defined %}
 81   [mon.{{ hostvars[host]['ansible_hostname'] }}]
 82     host = {{ hostvars[host]['ansible_hostname'] }}
 83     mon addr = {{ hostvars[host]['ansible_' + monitor_interface]['ipv4']['address'] }}
 84   {% endif %}
 85 {% endfor %}

• etcd
• consul
• look into better customization of docker unit for our needs

To be added later when socket activation code is available:

• volplugin
• netplugin

This issue track the discussion in #7 around evaluating a need for rearranging playbooks to provide for on-demand (or environment specific) installation of packages.

Just some helpers to:

• cd /opt/golang/src/github.com/contiv/... for netplugin/volplugin/etc
• source /etc/profile.d/envvar.sh
• start volplugin suite, netplugin suite
• start state stores
• Maybe go get some tooling necessary for both build systems?

Right now we can get clock skews during make or go build processes. I also worry this will cause desynchronization for etcd and consul.

The Docker apt repository seems to run into issues from time to time.

One solution to this problem would be to cache the apt lists and the deb package to allow the provisioning to continue using the cached deb package and the existing apt lists.

We're on Ubuntu 15.04 right now and that's an issue. Support is no longer provided for Ubuntu 15.04 and there's no point in tracking the latest unstable.

Perhaps we should switch to 14.04.

The current etcd playbooks fail occasionally if multiple nodes are configured at once (even after configuring the first node independently). I think it is due to a race condition with multple node trying to register membership with the master node at once.

This issue tracks debugging the condition further and fixing it in etcd playbook