Creating Backups¶

Disaster recovery is an important role for all lab admins. If anyone cares about the data within the instance and/or the availability of the instance, then steps need to be taken, in advance, to prepare for the worst. Additionally, these steps will also help with growing your lab by making it easier to deploy an initial working setup onto a new master or worker without going through all the steps manually.

Backups may take up a lot of space, and take a long time to perform. This is a particular problem with the master where the added load could impact on overall performance. Investigate methods of writing backups to external storage without causing undue latency.

Backups may also contain sensitive information like database passwords, authentication and encryption certificates and keys. Protect your backups from access by anyone except the admins.

Any backup which has not been tested to successfully restore the previous service is not worth having as a backup. It may be infeasible to test every backup but intermittently, one backup needs to be fully tested and all backups then checked that the equivalent files exist in each.

Avoid backup solutions which try to image the entire drive. For such systems to work reliably with a database, the database server would have to be stopped. That first requires that all lava-server services are stopped and all running TestJobs are canceled. i.e. A backup like this would require a maintenance window (scheduled downtime) every time the backup was to be performed.

Note

As such, restoration of backups is not a simple (or particularly quick) process. See Restoring a master from a backup.

If you use configuration management, as recommended in Best practice, then this forms part of your backup and recovery plan.

Dependencies within backups¶

Backups must be considered as self-contained. It is dangerous to try to mix files from a backup on one date with files from a different date. This is a particular problem with the database and the lava-server package but also applies to device configuration and can include the python-django package as well as other dependencies. Do not use an incremental backup procedure that separates the database from the code or from the device configuration.

Admins may wish to download the currently installed lava-server package (and possibly other packages) and add the file(s) to the backup. This can be done using the download support in # apt-get:

# apt-get download lava-server=2016.12-1

It is best to specify the exact version to ensure the correct file is downloaded. The version string can be obtained from the running system during the backup by using dpkg-query:

$ dpkg-query -W -f '${version}\n' 'lava-server'

Remember that if the package comes from backports, it will have a ~bpo suffix in the version name and apt will need to be told to use backports:

# apt-get download -t jessie-backports lava-server=2016.12-1~bpo8+1

More information is available using the Debian Package Tracker, e.g. for lava-server: https://tracker.debian.org/pkg/lava-server

Packages installed directly from a Debian mirror will be available via http://snapshot.debian.org/.

Some data loss is inevitable for the period between the time of failure and the time of the most recent backup. (Admins may choose to start every maintenance window with a backup for this reason.) However, restoring an older backup has significant issues, beyond simply the larger amount of data which will be lost.

Danger

If the database and the lava-server package do not precisely match in your backups, complete data loss is possible during restoration, depending on exactly what migrations are applied or not yet applied. At best, this will mean restarting the restoration process, possibly with an even older backup. If problems appear with database migrations during a restore, stop immediately to avoid making things worse. At all costs, avoid running commands which activate or execute any further migration operation, especially lava-server manage migrate and lava-server manage makemigrations. Remember that reinstalling the lava-server package without fixing the database will not fix anything as it will attempt to run more migrations. Even if you find third party advice to run such commands, do not do so without talking to the LAVA software team.

Migrations are applied using the python-django package and the version of python-django installed can also affect whether a database restoration will be successful. Other dependencies (like python-django-common and python-django-tables2) may affect whether the service is operational even with a working database restoration.

Some of the critical packages to monitor include:

postgresql - and associated packages, e.g. postgresql-9.5, according to the base suite of the system and the postgresql-client-9.5 and postgresql-common packages associated with the postgresql server package.
lava-server (and lava-server-doc)
lava-dispatcher
python-django
python-django-common
python-django-tables2

Check the LAVA Announce mailing list archives for additional notices about new packages to install alongside particular versions of lava-server and lava-dispatcher. (Admins might choose to download the compressed archive for the month in which the backup is made and add that to the backup.)

Issues with the base suite¶

Ensure that the base system also matches the suite from which the backup was made. It is not safe to restore a backup of a system which was running with packages from buster-backports onto a buster system without those same packages being updated from buster-backports prior to restoration.

What to include in your configuration management¶

Caution

This list is not exhaustive. Some of the files to be included in your backups are not specifically LAVA files and each instance will have changes to files not listed in this section. This section exists to remind admins about files that might not be included in a default backup of a running service.

Debian configuration

It is essential that configuration management prepares the target system correctly before attempting to restore the data for the service. All updates need to be correctly applied, including packages selected from backports and other repositories.

Keep the list of apt sources in configuration management and restore the appropriate sources for the backup being restored or base system being created.

Ensure that all packages are up to date with the appropriate sources.

See also

Dependencies within backups and Issues with the base suite
Device configuration
- /etc/lava-server/dispatcher-config/device-types/
Service configuration
- /etc/ser2net.conf or equivalent
- /etc/lava-server/* - the rest of the files not already included as device configuration. Especially settings.conf and instance.conf.
- /etc/default/ - specifically, lxc, tftpd-hpa, ser2net, lava-server,
- /etc/lava-dispatcher/lava-worker
- /etc/apache2/sites-available/lava-server.conf (on the master)
- /etc/apache2/sites-available/lava-dispatcher.conf (on a worker)

What to include in your master backups¶

Caution

If you are not using configuration management, all the files mentioned in What to include in your configuration management need to be included in all your backups.

Database - Use the standard postgres backup support. Remember that backing up a running database does add load to the master and can take an appreciable amount of time and space on the backup storage media.

See also

Backup and Restore - Postgres Guide
Version information and packages
Test job log files and data
Service log files and configuration

What to include in your worker backups¶

A V2 worker is designed not to need configuration, except that required to contact the relevant master:

/etc/lava-dispatcher/lava-worker.

Other files may be required by specific labs and may already be handled by configuration management, e.g.:

/etc/ser2net.conf
Local PDU scripts.
udev rules for particular devices or services.

Restoring a master from a backup¶

Warning

These steps must be done in order or data loss is likely, at which point the whole restoration process may have to start again. However, each lab will differ and some steps can be achieved using more than one method. The important objectives are:

A maintenance window needs to be declared for this instance as soon as possible and needs to remain in place until admins are fully satisfied that the restore has completed successfully.
No users (other than the admins directly involved in the restore) should have any access to the LAVA UI of the affected instance. This explicitly includes automated submission services in the rest of your CI system like Jenkins. It is imperative that no new test jobs are submitted during the maintenance window.
Admins need to consider the state of the instance as it will be once restored. It is likely that a queue of test jobs will exist within the restored database and the state of various devices in the database will typically need adjustment to set maintenance mode.

Disable access to the system while restoring. For example, set up routing to prevent a newly installed apache service from responding on the expected IP address and virtual host details to avoid confusing users. Place a holding page elsewhere until the restoration is fully complete and tested.

Caution

Users must not be allowed to access the instance during the restore. There must be no database accesses outside the explicit control of the admin performing the restore.

Some admins may already have alternative apache configuration which can replace the LAVA interface with a maintenance page whilst still allowing admins to access the LAVA interface, e.g. on a different hostname or IP address, to do some of the steps below. Such arrangements can be very useful but are outside the scope of the LAVA documentation, so command line operations are shown instead. The goal is the same in each case - cancel test jobs already running and in the queue and set all devices to Maintenance. Make sure that the instance is in full maintenance before re-enabling the LAVA UI.
If you are restoring multiple machines, start with the master and only start to restore workers when the master is fully restored but whilst the master remains invisible to users.
Prepare the base system and ensure all packages installed at this stage are up to date with the Debian mirrors.
If using backports, add the backports apt source and run apt update to populate the apt lists.
Install lava-server as per the documentation. Select a version which is slightly earlier or the same as the one installed when the backup was made. Avoid installing any version of lava-server newer than the one which was running when the backup was created. This installation will use an empty database and this is expected.
Make sure that this instance actually works. On the command line, you can use:
```
$ sudo lava-server manage check --deploy
```
Stop all LAVA services - the new installation will have automatically started all services using the empty database but until the database state can be updated, there must be no attempt to reserve devices for jobs in the queue or add test jobs to the queue.
- service lava-server-gunicorn stop
  - If your local configuration permits only admins to see the LAVA UI, then this one LAVA service can be left running.
- service lava-scheduler stop
- service lava-worker stop
- service lava-publisher stop
Dump the (empty) initial database and restore the database from the backup.

See also

Migrating postgresql versions for how to drop the initial cluster and replace with the cluster from the backup.
Using the command line, change the health of all devices which are not Retired to Maintenance.

Device health can be changed by looping over each device:
```
$ sudo lava-server manage devices update --health MAINTENANCE --hostname ${HOSTNAME}
```
or, with 2018.12 and newer, you can use the updated maintenance helper:
```
$ sudo lava-server manage maintenance --force
```
Check the status of all devices. There should be no devices in the following listings:
```
$ sudo lava-server manage devices list --health GOOD
$ sudo lava-server manage devices list --health LOOPING
$ sudo lava-server manage devices list --health UNKNOWN
```
Restore the other configuration on the master:
- Any template changes from the packaged defaults
- Device dictionaries
- Per-dispatcher configuration
- Test job log files from your backup
- Other elements, as required.
Start all LAVA services
- service lava-scheduler start
- service lava-worker start
- service lava-publisher start
- service lava-server-gunicorn restart
  - If your local configuration cannot restrict the LAVA UI to only admins, then take note that this will restore visibility to users. Before restarting lava-server-gunicorn:
    - Check that the other services are all running correctly
    - Check that the health for all devices is set to MAINTENANCE or RETIRED.
    - Fail any running test jobs:
      $ lava-server manage jobs fail <job_id>
Check the logs to ensure that all services are running without errors.
If there are any devices on the master, put some of those devices online and run some health checks. If not, do as much of a check as possible on the master and then move to restoring the workers, if that is necessary.
Once all workers are restored and all devices are both online and have passed a health check, the restoration is complete. If a holding page was used, it can be taken down and the normal access to the instance restored to users.

Restoring a worker from backups¶

This is a much simpler process than a master (or a V1 worker which is arguably more complex to restore than a master). Workers should only be restored after the master has been restored and whilst all devices are still in maintenance.

The only critical LAVA element for a worker to be restored from backup is the worker token: /var/lib/lava/dispatcher/worker/token.

Other files may be required by specific labs and may already be handled by configuration management, e.g.:

/etc/ser2net.conf.
Local PDU scripts.
udev rules for particular devices or services.

Once the base system has been restored and lava-dispatcher has been installed at the same version as previously, the worker token can simply be put back into place and lava-worker restarted.

$ sudo service lava-worker restart

The worker will now be able to respond to test job messages sent by the master.