Appendix F2: Series upgrade OpenStack¶
Overview¶
This document will provide specific steps for how to perform a series upgrade across the entirety of a Charmed OpenStack cloud.
Warning
This document is based upon the foundational knowledge and guidelines set forth in the more general Series upgrade appendix. That reference must be studied in-depth prior to attempting the steps outlined here. In particular, ensure that the Pre-upgrade requirements are satisfied; the Specific series upgrade procedures have been reviewed and considered; and that Workload specific preparations have been addressed during planning.
Downtime¶
Although the goal is to minimise downtime the series upgrade process across a cloud will nonetheless result in some level of downtime for the control plane.
When the machines associated with stateful applications such as percona-cluster and rabbitmq-server undergo a series upgrade all cloud APIs will experience downtime, in addition to the stateful application itself.
When machines associated with a single API application undergo a series upgrade that individual API will also experience downtime. This is because it is necessary to pause services in order to avoid race condition errors.
For those applications working in tandem with hacluster, as will be shown, some hacluster units will need to be paused before the upgrade. One should assume that the commencement of an outage coincides with this step (it will cause cluster quorum heartbeats to fail and the service VIP will consequently go offline).
Reference cloud topology¶
This section describes a hyperconverged cloud topology that this document will use for the procedural steps to follow. Hyperconvergence refers to the practice of co-locating principal applications on the same machine.
The topology is defined in this way:
Only compute and storage charms (and their subordinates) may be co-located.
Third-party charms either do not exist or have been thoroughly tested for a series upgrade.
The following are containerised:
All API applications
The percona-cluster application
The rabbitmq-server application
The ceph-mon application
Storage charms are charms that manage physical disks. For example, ceph-osd and swift-storage. Example OpenStack subordinate charms are networking SDN charms for the nova-compute charm, or monitoring charms for compute or storage charms.
Caution
If your cloud differs from this topology you must adapt the procedural steps accordingly. In particular, look at the aspects of co-located applications and containerised applications. Recall that:
the upgrade-series command:
affects all applications residing on the target machine
does not affect containers hosted on the target machine
an application’s leader should be upgraded before its non-leaders
Generalised OpenStack series upgrade¶
This section will summarise the series upgrade steps in the context of specific OpenStack applications. It is an enhancement of the Generic series upgrade section in the companion document.
Applications for which this summary does not apply include:
nova-compute
ceph-mon
ceph-osd
This is because the above applications do not require the pausing of units and application leadership is irrelevant for them.
However, this summary does apply to all API applications (e.g. neutron-api, keystone, nova-cloud-controller), as well as percona-cluster, and rabbitmq-server.
Important
The first machine to be upgraded is always associated with the leader of the principal application. Let this machine be called the “principal leader machine” and its unit be called the “principal leader unit”.
The steps are as follows:
Set the default series for the principal application and ensure the same has been done to the model.
If hacluster is used, pause the hacluster units not associated with the principal leader machine.
Pause the principal non-leader units.
Perform a series upgrade on the principal leader machine.
Perform any pre-upgrade workload maintenance tasks.
Invoke the prepare sub-command.
Upgrade the operating system (APT commands).
Perform any post-upgrade workload maintenance tasks.
Reboot.
Set the value of the (application-dependent)
openstack-origin
or thesource
configuration option to ‘distro’ (new operating system).Invoke the complete sub-command on the principal leader machine.
Repeat steps 4 and 6 for the application non-leader machines.
Perform any possible cluster completed upgrade tasks once all machines have had their series upgraded.
Note
Here is a non-extensive list of the most common post-upgrade tasks for OpenStack and supporting charms:
percona-cluster: run action
complete-cluster-series-upgrade
on the leader unit.rabbitmq-server: run action
complete-cluster-series-upgrade
on the leader unit.ceilometer: run action
ceilometer-upgrade
on the leader unit.vault: Each vault unit will need to be unsealed after its machine is rebooted.
Procedures¶
The procedures are categorised based on application types. The example scenario used throughout is a ‘xenial’ to ‘bionic’ series upgrade, within an OpenStack release of Queens (i.e. the starting point is a cloud archive pocket of ‘xenial-queens’).
Stateful applications¶
This section covers the series upgrade procedure for containerised stateful applications. These include:
ceph-mon
percona-cluster
rabbitmq-server
A stateful application is one that maintains the state of various aspects of the cloud. Clustered stateful applications, such as all the ones given above, also require a quorum to function properly. Because of these reasons a stateful application should not have all of its units restarted simultaneously; it must have the series of its corresponding machines upgraded sequentially.
Note
The concurrent upgrade approach is theoretically possible, although to use it all cloud workloads will need to be stopped in order to ensure consistency. This is not recommended.
The example procedure will be based on the percona-cluster application.
Important
Unlike percona-cluster, the ceph-mon and rabbitmq-server applications do not use hacluster to achieve HA, nor do they need backups. Disregard therefore the hacluster and backup steps for these two applications.
The ceph-mon charm will maintain the MON cluster during a series upgrade, so ceph-mon units do not need to be paused.
This scenario is represented by the following partial juju status command output:
Model Controller Cloud/Region Version SLA Timestamp
upgrade maas-controller mymaas/default 2.7.6 unsupported 18:26:57Z
App Version Status Scale Charm Store Rev OS Notes
percona-cluster 5.6.37 active 3 percona-cluster jujucharms 286 ubuntu
percona-cluster-hacluster active 3 hacluster jujucharms 66 ubuntu
Unit Workload Agent Machine Public address Ports Message
percona-cluster/0 active idle 0/lxd/0 10.0.0.47 3306/tcp Unit is ready
percona-cluster-hacluster/0* active idle 10.0.0.47 Unit is ready and clustered
percona-cluster/1* active idle 1/lxd/0 10.0.0.48 3306/tcp Unit is ready
percona-cluster-hacluster/2 active idle 10.0.0.48 Unit is ready and clustered
percona-cluster/2 active idle 2/lxd/0 10.0.0.49 3306/tcp Unit is ready
percona-cluster-hacluster/1 active idle 10.0.0.49 Unit is ready and clustered
In summary, the principal leader unit is percona-cluster/1 and is deployed on machine 1/lxd/0 (the principal leader machine).
Warning
During this upgrade, there will be a MySQL service outage. The HA resources provided by hacluster will not be monitored during the series upgrade due to the pausing of units.
Perform any workload maintenance pre-upgrade steps. For percona-cluster, take a backup and transfer it to a secure location:
juju run-action --wait percona-cluster/1 backup juju scp -- -r percona-cluster/1:/opt/backups/mysql /path/to/local/directory
Permissions will need to be altered on the remote machine, and note that the last command transfers all existing backups.
Set the default series for both the model and the principal application:
juju model-config default-series=bionic juju set-series percona-cluster bionic
Pause the hacluster units not associated with the principal leader machine:
juju run-action --wait percona-cluster-hacluster/0 pause juju run-action --wait percona-cluster-hacluster/1 pause
Pause the principal non-leader units:
juju run-action --wait percona-cluster/0 pause juju run-action --wait percona-cluster/2 pause
For percona-cluster, leaving the principal leader unit up will ensure it has the latest MySQL sequence number; it will be considered the most up to date cluster member.
Perform a series upgrade on the principal leader machine:
juju upgrade-series 1/lxd/0 prepare bionic juju run --machine=1/lxd/0 -- sudo apt update juju ssh 1/lxd/0 sudo apt full-upgrade juju ssh 1/lxd/0 sudo do-release-upgrade
For percona-cluster, there are no post-upgrade steps; the prompt to reboot can be answered in the affirmative.
Set the value of the
source
configuration option to ‘distro’:juju config percona-cluster source=distro
Invoke the complete sub-command on the principal leader machine:
juju upgrade-series 1/lxd/0 complete
At this point the juju status output looks like this:
Model Controller Cloud/Region Version SLA Timestamp upgrade maas-controller mymaas/default 2.7.6 unsupported 19:51:52Z App Version Status Scale Charm Store Rev OS Notes percona-cluster 5.7.20 maintenance 3 percona-cluster jujucharms 286 ubuntu percona-cluster-hacluster blocked 3 hacluster jujucharms 66 ubuntu Unit Workload Agent Machine Public address Ports Message percona-cluster/0 maintenance idle 0/lxd/0 10.0.0.47 3306/tcp Paused. Use 'resume' action to resume normal service. percona-cluster-hacluster/0* maintenance idle 10.0.0.47 Paused. Use 'resume' action to resume normal service. percona-cluster/1* active idle 1/lxd/0 10.0.0.48 3306/tcp Unit is ready percona-cluster-hacluster/2 blocked idle 10.0.0.48 Resource: res_mysql_11810cc_vip not running percona-cluster/2 maintenance idle 2/lxd/0 10.0.0.49 3306/tcp Paused. Use 'resume' action to resume normal service. percona-cluster-hacluster/1 maintenance idle 10.0.0.49 Paused. Use 'resume' action to resume normal service. Machine State DNS Inst id Series AZ Message 0 started 10.0.0.44 node1 xenial zone1 Deployed 0/lxd/0 started 10.0.0.47 juju-f83fcd-0-lxd-0 xenial zone1 Container started 1 started 10.0.0.45 node2 xenial zone2 Deployed 1/lxd/0 started 10.0.0.48 juju-f83fcd-1-lxd-0 bionic zone2 Running 2 started 10.0.0.46 node3 xenial zone3 Deployed 2/lxd/0 started 10.0.0.49 juju-f83fcd-2-lxd-0 xenial zone3 Container started
For percona-cluster, a sanity check should be done on the leader unit’s databases and data.
Repeat steps 5 and 7 for the principal non-leader machines.
Perform any possible cluster completed upgrade tasks once all machines have had their series upgraded:
juju run-action --wait percona-cluster/leader complete-cluster-series-upgrade
For percona-cluster (and rabbitmq-server), the above action is performed on the leader unit. It informs each cluster node that the upgrade process is complete cluster-wide. This also updates MySQL configuration with all peers in the cluster.
API applications¶
This section covers series upgrade procedures for containerised API applications. These include, but are not limited to:
cinder
glance
keystone
neutron-api
nova-cloud-controller
Machines hosting API applications can have their series upgraded concurrently because those applications are stateless. This results in a dramatically reduced downtime for the application. A sequential approach will not reduce downtime as the HA services will still need to be brought down during the upgrade associated with the application leader.
The following two sub-sections will show how to perform a series upgrade concurrently for a single API application and for multiple API applications.
Upgrading a single API application concurrently¶
This example procedure will be based on the keystone application.
This scenario is represented by the following partial juju status command output:
Model Controller Cloud/Region Version SLA Timestamp
upgrade maas-controller mymaas/default 2.7.6 unsupported 22:48:41Z
App Version Status Scale Charm Store Rev OS Notes
keystone 13.0.2 active 3 keystone jujucharms 312 ubuntu
keystone-hacluster active 3 hacluster jujucharms 66 ubuntu
Unit Workload Agent Machine Public address Ports Message
keystone/0* active idle 0/lxd/0 10.0.0.70 5000/tcp Unit is ready
keystone-hacluster/0* active idle 10.0.0.70 Unit is ready and clustered
keystone/1 active idle 1/lxd/0 10.0.0.71 5000/tcp Unit is ready
keystone-hacluster/2 active idle 10.0.0.71 Unit is ready and clustered
keystone/2 active idle 2/lxd/0 10.0.0.72 5000/tcp Unit is ready
keystone-hacluster/1 active idle 10.0.0.72 Unit is ready and clustered
In summary, the principal leader unit is keystone/0 and is deployed on machine 0/lxd/0 (the principal leader machine).
Set the default series for both the model and the principal application:
juju model-config default-series=bionic juju set-series keystone bionic
Pause the hacluster units not associated with the principal leader machine:
juju run-action --wait keystone-hacluster/1 pause juju run-action --wait keystone-hacluster/2 pause
Pause the principal non-leader units:
juju run-action --wait keystone/1 pause juju run-action --wait keystone/2 pause
Perform any workload maintenance pre-upgrade steps on all machines. There are no keystone-specific steps to perform.
Invoke the prepare sub-command on all machines, starting with the principal leader machine:
juju upgrade-series 0/lxd/0 prepare bionic juju upgrade-series 1/lxd/0 prepare bionic juju upgrade-series 2/lxd/0 prepare bionic
At this point the juju status output looks like this:
Model Controller Cloud/Region Version SLA Timestamp upgrade maas-controller mymaas/default 2.7.6 unsupported 23:11:01Z App Version Status Scale Charm Store Rev OS Notes keystone 13.0.2 blocked 3 keystone jujucharms 312 ubuntu keystone-hacluster blocked 3 hacluster jujucharms 66 ubuntu Unit Workload Agent Machine Public address Ports Message keystone/0* blocked idle 0/lxd/0 10.0.0.70 5000/tcp Ready for do-release-upgrade and reboot. Set complete when finished. keystone-hacluster/0* blocked idle 10.0.0.70 Ready for do-release-upgrade. Set complete when finished keystone/1 blocked idle 1/lxd/0 10.0.0.71 5000/tcp Ready for do-release-upgrade and reboot. Set complete when finished. keystone-hacluster/2 blocked idle 10.0.0.71 Ready for do-release-upgrade. Set complete when finished keystone/2 blocked idle 2/lxd/0 10.0.0.72 5000/tcp Ready for do-release-upgrade and reboot. Set complete when finished. keystone-hacluster/1 blocked idle 10.0.0.72 Ready for do-release-upgrade. Set complete when finished
Upgrade the operating system on all machines. The non-interactive method is used here:
juju run --machine=0/lxd/0,1/lxd/0,2/lxd/0 --timeout=10m \ -- sudo apt-get update juju run --machine=0/lxd/0,1/lxd/0,2/lxd/0 --timeout=60m \ -- sudo DEBIAN_FRONTEND=noninteractive apt-get --assume-yes \ -o "Dpkg::Options::=--force-confdef" \ -o "Dpkg::Options::=--force-confold" dist-upgrade juju run --machine=0/lxd/0,1/lxd/0,2/lxd/0 --timeout=120m \ -- sudo DEBIAN_FRONTEND=noninteractive \ do-release-upgrade -f DistUpgradeViewNonInteractive
Important
Choose values for the
--timeout
option that are appropriate for the task at hand.Perform any workload maintenance post-upgrade steps on all machines. There are no keystone-specific steps to perform.
Reboot all machines:
juju run --machine=0/lxd/0,1/lxd/0,2/lxd/0 -- sudo reboot
Set the value of the
openstack-origin
configuration option to ‘distro’:juju config keystone openstack-origin=distro
Invoke the complete sub-command on all machines:
juju upgrade-series 0/lxd/0 complete juju upgrade-series 1/lxd/0 complete juju upgrade-series 2/lxd/0 complete
Upgrading multiple API applications concurrently¶
This example procedure will be based on the nova-cloud-controller and glance applications.
This scenario is represented by the following partial juju status command output:
Model Controller Cloud/Region Version SLA Timestamp
upgrade maas-controller mymaas/default 2.7.6 unsupported 19:23:41Z
App Version Status Scale Charm Store Rev OS Notes
glance 16.0.1 active 3 glance jujucharms 295 ubuntu
glance-hacluster active 3 hacluster jujucharms 66 ubuntu
nova-cc-hacluster active 3 hacluster jujucharms 66 ubuntu
nova-cloud-controller 17.0.12 active 3 nova-cloud-controller jujucharms 343 ubuntu
Unit Workload Agent Machine Public address Ports Message
glance/0* active idle 0/lxd/0 10.246.114.39 9292/tcp Unit is ready
glance-hacluster/0* active idle 10.246.114.39 Unit is ready and clustered
glance/1 active idle 1/lxd/0 10.246.114.40 9292/tcp Unit is ready
glance-hacluster/1 active idle 10.246.114.40 Unit is ready and clustered
glance/2 active idle 2/lxd/0 10.246.114.41 9292/tcp Unit is ready
glance-hacluster/2 active idle 10.246.114.41 Unit is ready and clustered
nova-cloud-controller/0 active idle 3/lxd/0 10.246.114.48 8774/tcp,8778/tcp Unit is ready
nova-cc-hacluster/2 active idle 10.246.114.48 Unit is ready and clustered
nova-cloud-controller/1* active idle 4/lxd/0 10.246.114.43 8774/tcp,8778/tcp Unit is ready
nova-cc-hacluster/0* active idle 10.246.114.43 Unit is ready and clustered
nova-cloud-controller/2 active idle 5/lxd/0 10.246.114.47 8774/tcp,8778/tcp Unit is ready
nova-cc-hacluster/1 active idle 10.246.114.47 Unit is ready and clustered
In summary,
The glance principal leader unit is glance/0 and is deployed on machine 0/lxd/0 (the glance principal leader machine).
The nova-cloud-controller principal leader unit is nova-cloud-controller/1 and is deployed on machine 4/lxd/0 (the nova-cloud-controller principal leader machine).
The procedure has been expedited slightly by adding the --yes
confirmation
option to the prepare sub-command.
Set the default series for both the model and the principal applications:
juju model-config default-series=bionic juju set-series glance bionic juju set-series nova-cloud-controller bionic
Pause the hacluster units not associated with their principal leader machines:
juju run-action --wait glance-hacluster/1 pause juju run-action --wait glance-hacluster/2 pause juju run-action --wait nova-cc-hacluster/1 pause juju run-action --wait nova-cc-hacluster/2 pause
Pause the principal non-leader units:
juju run-action --wait glance/1 pause juju run-action --wait glance/2 pause juju run-action --wait nova-cloud-controller/0 pause juju run-action --wait nova-cloud-controller/2 pause
Perform any workload maintenance pre-upgrade steps on all machines. There are no glance-specific or nova-cloud-controller-specific steps to perform.
Invoke the prepare sub-command on all machines, starting with the principal leader machines:
juju upgrade-series --yes 0/lxd/0 prepare bionic juju upgrade-series --yes 4/lxd/0 prepare bionic juju upgrade-series --yes 1/lxd/0 prepare bionic juju upgrade-series --yes 2/lxd/0 prepare bionic juju upgrade-series --yes 3/lxd/0 prepare bionic juju upgrade-series --yes 5/lxd/0 prepare bionic
Upgrade the operating system on all machines. The non-interactive method is used here:
juju run --machine=0/lxd/0,1/lxd/0,2/lxd/0,3/lxd/0,4/lxd/0,5/lxd/0 \ --timeout=20m -- sudo apt-get update juju run --machine=0/lxd/0,1/lxd/0,2/lxd/0,3/lxd/0,4/lxd/0,5/lxd/0 \ --timeout=120m -- sudo DEBIAN_FRONTEND=noninteractive apt-get --assume-yes \ -o "Dpkg::Options::=--force-confdef" \ -o "Dpkg::Options::=--force-confold" dist-upgrade juju run --machine=0/lxd/0,1/lxd/0,2/lxd/0,3/lxd/0,4/lxd/0,5/lxd/0 \ --timeout=200m -- sudo DEBIAN_FRONTEND=noninteractive \ do-release-upgrade -f DistUpgradeViewNonInteractive
Perform any workload maintenance post-upgrade steps on all machines. There are no glance-specific or nova-cloud-controller-specific steps to perform.
Reboot all machines:
juju run --machine=0/lxd/0,1/lxd/0,2/lxd/0,3/lxd/0,4/lxd/0,5/lxd/0 -- sudo reboot
Set the value of the
openstack-origin
configuration option to ‘distro’:juju config glance openstack-origin=distro juju config nova-cloud-controller openstack-origin=distro
Invoke the complete sub-command on all machines:
juju upgrade-series 0/lxd/0 complete juju upgrade-series 1/lxd/0 complete juju upgrade-series 2/lxd/0 complete juju upgrade-series 3/lxd/0 complete juju upgrade-series 4/lxd/0 complete juju upgrade-series 5/lxd/0 complete
Physical machines¶
This section covers series upgrade procedures for applications hosted on physical machines in particular. These typically include:
ceph-osd
neutron-gateway
nova-compute
When performing a series upgrade on a physical machine more attention should be given to any workload maintenance pre-upgrade steps:
For compute nodes migrate all running VMs to another hypervisor.
For network nodes force HA routers off of the current node.
Any storage related tasks that may be required.
Any site specific tasks that may be required.
The following two sub-sections will show how to perform a series upgrade for a single physical machine and for multiple physical machines concurrently.
Upgrading a single physical machine¶
This example procedure will be based on the nova-compute and ceph-osd applications residing on the same physical machine. Since application leadership does not play a significant role with these two applications, and because the hacluster application is not present, there will be no units to pause (as there were in previous scenarios).
This scenario is represented by the following partial juju status command output:
Model Controller Cloud/Region Version SLA Timestamp
upgrade maas-controller mymaas/default 2.7.6 unsupported 15:23:21Z
App Version Status Scale Charm Store Rev OS Notes
ceph-osd 12.2.12 active 1 ceph-osd jujucharms 301 ubuntu
keystone 13.0.2 active 1 keystone jujucharms 312 ubuntu
nova-compute 17.0.12 active 1 nova-compute jujucharms 314 ubuntu
Unit Workload Agent Machine Public address Ports Message
ceph-osd/0* active idle 0 10.0.0.235 Unit is ready (1 OSD)
keystone/0* active idle 0/lxd/0 10.0.0.240 5000/tcp Unit is ready
nova-compute/0* active idle 0 10.0.0.235 Unit is ready
Machine State DNS Inst id Series AZ Message
0 started 10.0.0.235 node1 xenial zone1 Deployed
0/lxd/0 started 10.0.0.240 juju-88b27a-0-lxd-0 xenial zone1 Container started
In summary, the ceph-osd and nova-compute applications are hosted on machine 0. Recall that container 0/lxd/0 will need to have its series upgraded separately.
It is recommended to set the Ceph cluster OSDs to ‘noout’. This is typically done at the application level (i.e. not at the unit or machine level):
juju run-action --wait ceph-mon/leader set-noout
All running VMs should be migrated to another hypervisor.
Upgrade the series on machine 0:
Invoke the prepare sub-command:
juju upgrade-series 0 prepare bionic
Upgrade the operating system:
juju run --machine=0 -- sudo apt update juju ssh 0 sudo apt full-upgrade juju ssh 0 sudo do-release-upgrade
Reboot (if not already done):
juju run --machine=0 -- sudo reboot
Set the value of the
openstack-origin
orsource
configuration options to ‘distro’:juju config nova-compute openstack-origin=distro juju config ceph-osd source=distro
Invoke the complete sub-command on the machine:
juju upgrade-series 0 complete
If OSDs were previously set to ‘noout’ then check up/in status of those OSDs in ceph status, then unset ‘noout’ for the cluster:
juju run --unit ceph-mon/leader -- ceph status juju run-action --wait ceph-mon/leader unset-noout
Upgrading multiple physical hosts concurrently¶
When physical machines have their series upgraded concurrently Availability Zones need to be taken into account. Machines should be placed into upgrade groups such that any API services running on them have a maximum of one unit per group. This is to ensure API availability at the reboot stage.
This simplified bundle is used to demonstrate the general idea:
series: xenial
machines:
0: {}
1: {}
2: {}
3: {}
4: {}
5: {}
applications:
nova-compute:
charm: cs:nova-compute
num_units: 3
options:
openstack-origin: cloud:xenial-queens
to:
- 0
- 2
- 4
keystone:
charm: cs:keystone
constraints: mem=1G
num_units: 3
options:
vip: 10.85.132.200
openstack-origin: cloud:xenial-queens
to:
- lxd:1
- lxd:3
- lxd:5
keystone-hacluster:
charm: cs:hacluster
options:
cluster_count: 3
Three upgrade groups could consist of the following machines:
Machines 0 and 1
Machines 2 and 3
Machines 4 and 5
In this way, a less time-consuming series upgrade can be performed while still ensuring the availability of services.
Caution
For the ceph-osd application, ensure that rack-aware replication rules exist
in the CRUSH map if machines are being rebooted together. This is to prevent
significant interruption to running workloads from occurring if the
same placement group is hosted on those machines. For example, if ceph-mon
is deployed with customize-failure-domain
set to ‘true’ and the ceph-osd
units are hosted on machines in three or more separate Juju AZs you can
safely reboot ceph-osd machines concurrently in the same zone. See
Ceph AZ in OpenStack high availability for
details.
Automation¶
Series upgrades across an OpenStack cloud can be time consuming, even when using concurrent methods wherever possible. They can also be tedious and thus susceptible to human error.
The following code examples encapsulate the processes described in this document. They are provided solely to illustrate the methods used to develop and test the series upgrade primitives:
Parallel tests: An example that is used as a functional verification of a series upgrade in the OpenStack Charms project.
Upgrade helpers: A set of helpers used in the above upgrade example.
Caution
The example code should only be used for its intended use case of development and testing. Do not attempt to automate a series upgrade on a production cloud.