OpenStack upgrade

This document outlines how to upgrade a Juju-deployed OpenStack cloud.

Warning

Upgrading an OpenStack cloud is not risk-free. The procedures outlined in this guide should first be tested in a pre-production environment.

Please read the Upgrades overview page before continuing.

Note

The charms only support single-step OpenStack upgrades (N+1). That is, to upgrade two releases forward you need to upgrade twice. You cannot skip releases when upgrading OpenStack with charms.

It may be worthwhile to read the upstream OpenStack Upgrades guide.

Release Notes

The OpenStack Charms Release Notes for the corresponding current and target versions of OpenStack must be consulted for any special instructions. In particular, pay attention to services and/or configuration options that may be retired, deprecated, or changed.

Manual intervention

It is intended that the now upgraded charms are able to accommodate all software changes associated with the corresponding OpenStack services to be upgraded. A new charm will also strive to produce a service as similarly configured to the pre-upgraded service as possible.

However, there are still times when intervention on the part of the operator may be needed, such as when an OpenStack service is removed/added/replaced or when a software bug (in the charms or in upstream OpenStack) affecting the upgrade is present. The below resources cover these topics:

Ensure cloud node software is up to date

Every machine in the cloud, including containers, should have their software packages updated to ensure that the latest SRUs have been applied. This is done in the usual manner:

sudo apt-get update
sudo apt-get dist-upgrade

Verify the current deployment

Confirm that the output for the juju status command of the current deployment is error-free. In addition, if monitoring is in use (e.g. Nagios), ensure that all alerts have been resolved. You may also consider running a battery of operational checks on the cloud.

This step is to make certain that any issues that may appear after the upgrade are not due to pre-existing problems.

Disable unattended-upgrades

When performing a service upgrade on a cloud node that hosts multiple principle charms (e.g. nova-compute and ceph-osd), ensure that unattended-upgrades is disabled on the underlying machine for the duration of the upgrade process. This is to prevent the other services from being upgraded outside of Juju’s control. On a cloud node run:

sudo dpkg-reconfigure -plow unattended-upgrades

Perform a database backup

Before making any changes to cloud services perform a backup of the cloud database by running the backup action on any single percona-cluster unit:

juju run-action --wait percona-cluster/0 backup

Now transfer the backup directory to the Juju client with the intention of subsequently storing it somewhere safe. This command will grab all existing backups:

juju scp -- -r percona-cluster/0:/opt/backups/mysql /path/to/local/directory

Permissions may first need to be altered on the remote machine.

Archive old database data

During the upgrade, database migrations will be run. This operation can be optimised by first archiving any stale data (e.g. deleted instances). Do this by running the archive-data action on any single nova-cloud-controller unit:

juju run-action --wait nova-cloud-controller/0 archive-data

This action may need to be run multiple times until the action output reports ‘Nothing was archived’.

Purge old compute service entries

Old compute service entries for units which are no longer part of the model should be purged before the upgrade. These entries will show as ‘down’ (and be hosted on machines no longer in the model) in the current list of compute services:

openstack compute service list

To remove a compute service:

openstack compute service delete <service-id>

Subordinate charm applications

Applications that are associated with subordinate charms are upgraded along with their parent application. Subordinate charms do not support the openstack-origin configuration option which, as will be shown, is a pre-requisite for initiating an OpenStack charm payload upgrade.

Upgrade order

Generally speaking, the order is determined by the idea of a dependency tree. Those services that have the most potential impact on other services are upgraded first and those services that have the least potential impact on other services are upgraded last.

In the below table charms are listed in the order in which their corresponding OpenStack services should be upgraded. Each service represented by a charm will need to be upgraded individually, and only the packages associated with a charm’s OpenStack service will be updated.

The order provided below is the order used by internal testing.

Order

Charm

1

rabbitmq-server

2

ceph-mon

3

keystone

4

aodh

5

barbican

6

ceilometer

7

ceph-fs

8

ceph-radosgw

9

cinder

10

designate

11

designate-bind

12

glance

13

gnocchi

14

heat

15

manila

16

manila-generic

17

neutron-api

18

neutron-gateway

19

placement

20

nova-cloud-controller

21

openstack-dashboard

22

nova-compute

23

ceph-osd

24

swift-proxy

25

swift-storage

26

octavia

Important

Services whose software is not included in the Ubuntu Cloud Archive are not represented in the above list. This software is upgraded by the administrator (on the units) using traditional means (e.g. manually via package tools or as part of a series upgrade). Common charms where this applies are ntp, memcached, percona-cluster, rabbitmq-server, mysql-innodb-cluster, and mysql-router.

Note

An Octavia upgrade may entail an update of its load balancers (amphorae) as a post-upgrade task. Reasons for doing this include:

  • API incompatibility between the amphora agent and the new Octavia service

  • the desire to use features available in the new amphora agent or haproxy

See the upstream documentation on Rotating amphora images.

Software sources

A key part of an OpenStack upgrade is the stipulation of a unit’s software sources. For an upgrade, the latter will naturally reflect a more recent combination of Ubuntu release (series) and OpenStack release. This combination is based on the Ubuntu Cloud Archive and translates to a “cloud archive OpenStack release”. It takes on the following syntax:

<ubuntu series>-<openstack-release>

The value is passed to a charm’s openstack-origin configuration option. For example, to select the ‘focal-victoria’ release:

openstack-origin=cloud:focal-victoria

In this way the charm is informed on where to find updates for the packages that it is responsible for.

Note

A few charms use option source instead of openstack-origin. See the next section.

Notes concerning the value of openstack-origin:

  • The default is ‘distro’. This denotes an Ubuntu release’s default archive (e.g. in the case of the focal series it corresponds to OpenStack Ussuri). The value of ‘distro’ is therefore invalid in the context of an OpenStack upgrade.

  • It should normally be the same across all charms.

  • Its series component must be that of the series currently in use (i.e. a series upgrade and an OpenStack upgrade are two completely separate procedures).

Perform the upgrade

There are three methods available for performing an OpenStack service upgrade. The appropriate method is chosen based on the actions supported by the charm. Actions for a charm can be listed with command juju actions <charm-name>.

All-in-one

The “all-in-one” method upgrades an application immediately. Although it is the quickest route, it can be harsh when applied in the context of multi-unit applications. This is because all the units are upgraded simultaneously, and is likely to cause a transient service outage. This method must be used if the application has a sole unit.

Attention

The “all-in-one” method should only be used when the charm does not support the openstack-upgrade action.

The syntax is:

juju config <openstack-charm> openstack-origin=cloud:<cloud-archive-release>

Charms whose services are not technically part of the OpenStack project will use the source charm option instead. The Ceph charms are a classic example:

juju config ceph-mon source=cloud:focal-victoria

Note

The ceph-osd and ceph-mon charms are able to maintain service availability during the upgrade.

So to upgrade Cinder across all units (currently running Focal) from Ussuri to Victoria:

juju config cinder openstack-origin=cloud:focal-victoria

Single-unit

The “single-unit” method builds upon the “all-in-one” method by allowing for the upgrade of individual units in a controlled manner. It requires the enablement of charm option action-managed-upgrade and the charm action openstack-upgrade.

Attention

The “single-unit” method should only be used when the charm does not support the pause and resume actions.

As a general rule, whenever there is the possibility of upgrading units individually, always upgrade the application leader first. The leader is the unit with a * next to it in the juju status output. It can also be discovered via the CLI:

juju run --application <application-name> is-leader

For example, to upgrade a three-unit glance application from Ussuri to Victoria where glance/1 is the leader:

juju config glance action-managed-upgrade=True
juju config glance openstack-origin=cloud:focal-victoria

juju run-action --wait glance/1 openstack-upgrade
juju run-action --wait glance/0 openstack-upgrade
juju run-action --wait glance/2 openstack-upgrade

Note

The openstack-upgrade action is only available for charms whose services are part of the OpenStack project. For instance, you will need to use the “all-in-one” method for the Ceph charms.

Paused-single-unit

The “paused-single-unit” method extends the “single-unit” method by allowing for the upgrade of individual units while paused. Additional charm requirements are the pause and resume actions. This method provides more versatility by allowing a unit to be removed from service, upgraded, and returned to service. Each of these are distinct events whose timing is chosen by the operator.

Attention

The “paused-single-unit” method is the recommended OpenStack service upgrade method.

For example, to upgrade a three-unit nova-compute application from Ussuri to Victoria where nova-compute/0 is the leader:

juju config nova-compute action-managed-upgrade=True
juju config nova-compute openstack-origin=cloud:focal-victoria

juju run-action --wait nova-compute/0 pause
juju run-action --wait nova-compute/0 openstack-upgrade
juju run-action --wait nova-compute/0 resume

juju run-action --wait nova-compute/1 pause
juju run-action --wait nova-compute/1 openstack-upgrade
juju run-action --wait nova-compute/1 resume

juju run-action --wait nova-compute/2 pause
juju run-action --wait nova-compute/2 openstack-upgrade
juju run-action --wait nova-compute/2 resume

In addition, this method also permits a possible hacluster subordinate unit, which typically manages a VIP, to be paused so that client traffic will not flow to the associated parent unit while its upgrade is underway.

Attention

When there is an hacluster subordinate unit then it is recommended to always take advantage of the “pause-single-unit” method’s ability to pause it before upgrading the parent unit.

For example, to upgrade a three-unit keystone application from Ussuri to Victoria where keystone/2 is the leader:

juju config keystone action-managed-upgrade=True
juju config keystone openstack-origin=cloud:focal-victoria

juju run-action --wait keystone-hacluster/1 pause
juju run-action --wait keystone/2 pause
juju run-action --wait keystone/2 openstack-upgrade
juju run-action --wait keystone/2 resume
juju run-action --wait keystone-hacluster/1 resume

juju run-action --wait keystone-hacluster/2 pause
juju run-action --wait keystone/1 pause
juju run-action --wait keystone/1 openstack-upgrade
juju run-action --wait keystone/1 resume
juju run-action --wait keystone-hacluster/2 resume

juju run-action --wait keystone-hacluster/0 pause
juju run-action --wait keystone/0 pause
juju run-action --wait keystone/0 openstack-upgrade
juju run-action --wait keystone/0 resume
juju run-action --wait keystone-hacluster/0 resume

Warning

The hacluster subordinate unit number may not necessarily match its parent unit number. As in the above example, only for keystone/0 do the unit numbers correspond (i.e. keystone-hacluster/0 is the subordinate unit).

Verify the new deployment

Check for errors in juju status output and any monitoring service.