Bare Metal Service Upgrade Guide¶

RPC version pinning and versioned object backports¶

Through careful RPC versioning, newer services are able to talk to older services (and vice-versa). The [DEFAULT]/pin_release_version configuration option is used for this. It should be set (pinned) to the release version that the older services are using. The newer services will backport RPC calls and objects to their appropriate versions from the pinned release. If the IncompatibleObjectVersion exception occurs, it is most likely due to an incorrect or unspecified [DEFAULT]/pin_release_version configuration value. For example, when [DEFAULT]/pin_release_version is not set to the older release version, no conversion will happen during the upgrade.

Online data migrations¶

To make database schema migrations less painful to execute, we have implemented process changes to facilitate upgrades.

• All data migrations are banned from schema migration scripts.
• Schema migration scripts only update the database schema.
• Data migrations must be done at the end of the rolling upgrade process, after the schema migration and after the services have been upgraded to the latest release.

All data migrations are performed using the ironic-dbsync online_data_migrations command. It can be run as a background process so that it does not interrupt running services; however it must be run to completion for a cold upgrade if the intent is to make use of new features immediately.

(You would also execute the same command with services turned off if you are doing a cold upgrade).

This data migration must be completed. If not, you will not be able to upgrade to future releases. For example, if you had upgraded from Ocata to Pike but did not do the data migrations, you will not be able to upgrade from Pike to Queens. (More precisely, you will not be able to apply Queens’ schema migrations.)

Graceful conductor service shutdown¶

The ironic-conductor service is a Python process listening for messages on a message queue. When the operator sends the SIGTERM signal to the process, the service stops consuming messages from the queue, so that no additional work is picked up. It completes any outstanding work and then terminates. During this process, messages can be left on the queue and will be processed after the Python process starts back up. This gives us a way to shutdown a service using older code, and start up a service using newer code with minimal impact.

Nodes that are being acted upon by an ironic-conductor process, which are not in a stable state, may encounter failures. Node failures that occur during an upgrade are likely due to timeouts, resulting from delays involving messages being processed and acted upon by a conductor during long running, multi-step processes such as deployment or cleaning.

API load balancer draining¶

If you are using a load balancer for the ironic-api services, we recommend that you redirect requests to the new API services and drain off of the ironic-api services that have not yet been upgraded.

Before maintenance window¶

• Upgrade the ironic-python-agent image

• Using the new release (ironic code), execute the required database schema updates by running the database upgrade command: ironic-dbsync upgrade. These schema change operations should have minimal or no effect on performance, and should not cause any operations to fail (but please check the release notes). You can:

  • install the new release on an existing system
  • install the new release in a new virtualenv or a container

  At this point, new columns and tables may exist in the database. These database schema changes are done in a way that both the old and new (N and N+1) releases can perform operations against the same schema.

During maintenance window¶

1. All ironic-conductor services should be upgraded first. Ensure that at least one ironic-conductor service is running at all times. For every ironic-conductor, either one by one or a few at a time:

   • shut down the service. Messages from the ironic-api services to the conductors are load-balanced by the message queue and a hash-ring, so the only thing you need to worry about is to shut the service down gracefully (using SIGTERM signal) to make sure it will finish all the requests being processed before shutting down.
   • upgrade the installed version of ironic and dependencies
   • set the [DEFAULT]/pin_release_version configuration option value to the version you are upgrading from (that is, the old version). Based on this setting, the new ironic-conductor services will downgrade any RPC communication and data objects to conform to the old service. For example, if you are upgrading from Ocata to Pike, set this value to ocata.
   • start the service

2. The next service to upgrade is ironic-api. Ensure that at least one ironic-api service is running at all times. You may want to start another temporary instance of the older ironic-api to handle the load while you are upgrading the original ironic-api services. For every ironic-api service, either one by one or a few at a time:

   • in HA deployment you are typically running them behind a load balancer (for example HAProxy), so you need to take the service instance out of the balancer
   • shut it down
   • upgrade the installed version of ironic and dependencies
   • set the [DEFAULT]/pin_release_version configuration option value to the version you are upgrading from (that is, the old version). Based on this setting, the new ironic-api services will downgrade any RPC communication and data objects to conform to the old service. For example, if you are upgrading from Ocata to Pike, set this value to ocata.
   • restart the service
   • add it back into the load balancer

   After upgrading all the ironic-api services, the Bare Metal service is running in the new version but with downgraded RPC communication and database object storage formats. New features can fail when objects are in the downgraded object formats and some internal RPC API functions may still not be available.

3. For all the ironic-conductor services, one at a time:

   • remove the [DEFAULT]/pin_release_version configuration option setting
   • restart the ironic-conductor service

4. For all the ironic-api services, one at a time:

   • remove the [DEFAULT]/pin_release_version configuration option setting
   • restart the ironic-api service

After maintenance window¶

Now that all the services are upgraded, the system is able to use the latest version of the RPC protocol and able to access all the features of the new release.

• Update any applicable configuration options to stop using any deprecated features or options, and perform any required work to transition to alternatives. All the deprecated features and options will be supported for one release cycle, so should be removed before your next upgrade is performed.

• Upgrade python-ironicclient along with other services connecting to the Bare Metal service as a client, such as nova-compute.

• Run the ironic-dbsync online_data_migrations command to make sure that data migrations are applied. The command lets you limit the impact of the data migrations with the --max-count option, which limits the number of migrations executed in one run. You should complete all of the migrations as soon as possible after the upgrade.

  Warning

  Note that you will not be able to start an upgrade to the next release after this one, until this has been completed for the current release. For example, as part of upgrading from Ocata to Pike, you need to complete Pike’s data migrations. If this not done, you will not be able to upgrade to Queens – it will not be possible to execute Queens’ database schema updates.