OpenStack upgrade

This document outlines how to upgrade the OpenStack service components of a Charmed 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.

Upgradable services

Only services whose software is included in the Ubuntu Cloud Archive will get upgraded during an OpenStack upgrade.

Services that are associated with subordinate charms are upgradable but only indirectly. They get upgraded along with their parent principal application.

Non-UCA software is upgraded by the administrator (on the units) using other means (e.g. manually via package utilities, the Landscape management tool, a snap, or as part of a series upgrade). Common applications where this applies are:

  • hacluster

  • memcached

  • mysql-innodb-cluster

  • mysql-router

  • ntp

  • percona-cluster

  • rabbitmq-server

  • vault

Prepare for the upgrade

Pay special attention to the below pre-upgrade preparatory and informational sections.

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

By design, the latest stable charms will support the software changes related to the OpenStack services being upgraded. During the upgrade, the charms will also strive to preserve the existing configuration of their associated services. Upstream OpenStack is also designed to support N+1 upgrades. However, there may still be times when intervention on the part of the operator is needed. The Issues, charm procedures, and OpenStack upgrade notes page covers this topic.

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 update
sudo apt full-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 are apparent after the upgrade are not due to pre-existing problems.

Perform pre-upgrade steps

Perform the pre-upgrade steps described in the below sections.

Disable unattended-upgrades

When performing a service upgrade on a cloud node that hosts multiple principal 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 backup of the service databases

Perform a backup of the cloud service databases by applying the mysqldump action to any unit of the cloud’s database application. Be sure to select all applicable databases; the commands provided are examples only.

The permissions on the remote backup directory will need to be adjusted in order to access the data. Take note that the transfer method presented here will capture all existing backups in that directory.

Important

Store the backup archive in a safe place.

The next two sections include the commands to run for the two possible database applications.

percona-cluster

The percona-cluster application requires a modification to its “strict mode” (see Percona strict mode for an understanding of the implications).

juju run-action --wait percona-cluster/0 set-pxc-strict-mode mode=MASTER
juju run-action --wait percona-cluster/0 mysqldump \
   databases=aodh,cinder,designate,glance,gnocchi,horizon,keystone,neutron,nova,nova_api,nova_cell0,placement
juju run-action --wait percona-cluster/0 set-pxc-strict-mode mode=ENFORCING

juju run -u percona-cluster/0 -- sudo chmod o+rx /var/backups/mysql
juju scp -- -r percona-cluster/0:/var/backups/mysql .
juju run -u percona-cluster/0 -- sudo chmod o-rx /var/backups/mysql

mysql-innodb-cluster

juju run-action --wait mysql-innodb-cluster/0 mysqldump \
   databases=cinder,designate,glance,gnocchi,horizon,keystone,neutron,nova,nova_api,nova_cell0,placement,vault

juju run -u mysql-innodb-cluster/0 -- sudo chmod o+rx /var/backups/mysql
juju scp -- -r mysql-innodb-cluster/0:/var/backups/mysql .
juju run -u mysql-innodb-cluster/0 -- sudo chmod o-rx /var/backups/mysql

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 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 prior to upgrading. 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>

List the upgrade order

Generally speaking, the upgrade 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. Note that since charms merely modify a machine’s apt sources, any co-located service will have their packages updated along with those of the service being targeted.

Warning

Ceph may require one of its options to be set prior to upgrading, and failure to consider this may result in a broken cluster. See the associated upgrade issue.

Note

Only stable charms are listed in the upgrade order table.

Order

Charm

1

ceph-mon

2

keystone

3

aodh

4

barbican

5

ceilometer

6

ceph-fs

7

ceph-radosgw

8

cinder

9

designate

10

designate-bind

11

glance

12

gnocchi

13

heat

14

manila

15

manila-ganesha

16

neutron-api

17

neutron-gateway or ovn-dedicated-chassis

18

ovn-central

19

placement

20

nova-cloud-controller

21

nova-compute

22

openstack-dashboard

23

ceph-osd

24

swift-proxy

25

swift-storage

26

octavia

Important

The OVN control plane will not be available between the commencement of the ovn-central upgrade and the completion of the nova-compute upgrade.

Perform the upgrade

There are three methods available for performing an OpenStack service upgrade, two of which have charm requirements in terms of supported actions. Each method also has advantages and disadvantages with regard to:

  • the time required to perform an upgrade

  • maintaining service availability during an upgrade

This table summarises the characteristics and requirements of each method:

Method

Time

Downtime

Charm requirements (actions)

all-in-one

shortest

most

none

single-unit

medium

medium

openstack-upgrade

paused-single-unit

longest

least

openstack-upgrade, pause, and resume

For example, although the all-in-one method upgrades a service the fastest, it also has the greatest potential for service downtime.

Note

A charm’s supported actions can be listed with command juju actions <application-name>.

As a general rule, whenever there is the possibility of upgrading units individually, always upgrade the application leader first.

Note

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 -a <application-name> is-leader

Depending on whether a given charm uses channels or not (see the Charm types page), there are slight differences in the upgrade procedures:

  • With non-channel charms, you will need to set the software sources of the charm. Guidance is provided in the Software sources page.

  • With channel charms, you will need to change the charm’s channel. See the Changing the channel section of the Charm delivery page for background information. Notably, a channel change will typically cause the underlying cloud service to restart.

All-in-one

The all-in-one method upgrades all application units simultaneously. This method must be used if the application has a sole unit.

Although it is the quickest route, it will also cause a temporary disruption of the corresponding service.

Important

Exceptionally, the ceph-osd and ceph-mon applications use the all-in-one method but their charms are able to maintain service availability during the upgrade.

For example, to upgrade Cinder across all units (currently running Focal) from Xena to Yoga:

juju config cinder action-managed-upgrade=False

If charm channels are in use:

juju refresh --channel yoga/stable cinder

If charm channels are not in use:

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

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. The charm must support the openstack-upgrade action, which in turn guarantees the availability of the action-managed-upgrade option.

This method is slower than the all-in-one method due to the need for each unit to be upgraded separately. There is a lesser chance of downtime as the unit being upgraded must be in the process of servicing client requests for downtime to occur.

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

juju config glance action-managed-upgrade=True

If charm channels are in use:

juju refresh --channel yoga/stable glance

If charm channels are not in use:

juju config glance openstack-origin=cloud:focal-yoga

In all cases, continue with the following commands (note that the leader has the openstack-upgrade action applied first):

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

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.

This is the slowest method due to the need for each unit to be upgraded separately in addition to the required pause/resume management. However, it is the method that will result in the least downtime as clients will not be able to solicit a paused service.

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

juju config nova-compute action-managed-upgrade=True

If charm channels are in use:

juju refresh --channel yoga/stable nova-compute

If charm channels are not in use:

juju config nova-compute openstack-origin=cloud:focal-yoga

In all cases, continue with the following commands (note that the leader has the openstack-upgrade action applied first):

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

Paused-single-unit with hacluster

In addition, this method also permits a possible hacluster subordinate unit, which typically manages a VIP, to be paused so that client requests will never even be directed to the associated parent unit.

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 Xena to Yoga where keystone/2 is the leader:

juju config keystone action-managed-upgrade=True

If charm channels are in use:

juju refresh --channel yoga/stable keystone

If charm channels are not in use:

juju config keystone openstack-origin=cloud:focal-yoga

Recall that the hacluster charm does not represent software found in the UCA. Its channel is therefore not changed here.

In all cases, continue with the following commands (note that the leader has the openstack-upgrade action applied first):

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 its subordinate unit).

Re-enable unattended-upgrades

In a previous step, unattended-upgrades were disabled on those cloud nodes that hosted multiple principal charms. Once such a node has had all of its services upgraded, unattended-upgrades should be re-enabled:

sudo dpkg-reconfigure -plow unattended-upgrades

Verify the new deployment

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

Example upgrade

The OpenStack upgrade example page shows the explicit steps used to upgrade a basic cloud.

Automation

Charmed OpenStack Upgrader (COU) is a tool that standardizes and automates upgrade procedures. COU encapsulates the processes described in this document and on the Charms upgrade page, which remain the primary references for performing OpenStack (payload) and charm upgrades respectively.

Important

Charmed OpenStack Upgrader should be tested thoroughly before using it in production. See the COU documentation for information on its usage and level of readiness.