Upgrades¶

RPC version pinning¶

Through careful RPC versioning, newer services are able to talk to older services (and vice-versa). The versions are autodetected using information reported in services table. In case of receiving CappedVersionUnknown or ServiceTooOld exceptions on service start, you’re probably having some old orphaned records in that table.

Graceful service shutdown¶

Many cinder services are python processes listening for messages on an AMQP queue. When the operator sends SIGTERM signal to the process, it stops getting new work from its queue, completes any outstanding work and then terminates. During this process, messages can be left on the queue for when 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.

Database upgrades¶

Cinder has two types of database upgrades in use:

• Schema migrations

• Data migrations

Schema migrations are defined in cinder/db/migrations/versions. They are the routines that transform our database structure, which should be additive and able to be applied to a running system before service code has been upgraded.

Data migrations are banned from schema migration scripts and are instead defined in cinder/db/api.py. They are kept separate to make DB schema migrations less painful to execute. Instead, the migrations are executed by a background process in a manner that doesn’t interrupt running services (you can also execute online data migrations with services turned off if you’re doing a cold upgrade). The cinder-manage db online_data_migrations utility can be used for this purpose. Before upgrading N to N+1, you need to run this tool in the background until it tells you no more migrations are needed. Note that you won’t be able to apply N+1’s schema migrations before completing N’s online data migrations.

For information on developing your own schema or data migrations as part of a feature or bugfix, refer to Database migrations.

API load balancer draining¶

When upgrading API nodes, you can make your load balancer only send new connections to the newer API nodes, allowing for a seamless update of your API nodes.

DB prune deleted rows¶

Currently resources are soft deleted in the database, so users are able to track instances in the DB that are created and destroyed in production. However, most people have a data retention policy, of say 30 days or 90 days after which they will want to delete those entries. Not deleting those entries affects DB performance as indices grow very large and data migrations take longer as there is more data to migrate. To make pruning easier there’s a cinder-manage db purge <age_in_days> command that permanently deletes records older than specified age.

Versioned object backports¶

RPC pinning ensures new services can talk to the older service’s method signatures. But many of the parameters are objects that may well be too new for the old service to understand. Cinder makes sure to backport an object to a version that it is pinned to before sending.

Plan your upgrade¶

• Read and ensure you understand the release notes for the next release.

• Make a backup of your database. Cinder does not support downgrading of the database. Hence, in case of upgrade failure, restoring database from backup is the only choice.

• To avoid dependency hell it is advised to have your Cinder services deployed separately in containers or Python venvs.

  Note

  Cinder is basing version detection on what is reported in the services table in the DB. Before upgrade make sure you don’t have any orphaned old records there, because these can block starting newer services. You can clean them up using cinder-manage service remove <binary> <host> command.

Note that there’s an assumption that live upgrade can be performed only between subsequent releases. This means that you cannot upgrade N directly to N+2, you need to upgrade to N+1 first.

The assumed service upgrade order is cinder-scheduler, cinder-volume, cinder-backup and finally cinder-api.

Rolling upgrade process¶

To reduce downtime, the services can be upgraded in a rolling fashion. It means upgrading a few services at a time. To minimise downtime you need to have HA Cinder deployment, so at the moment a service is upgraded, you’ll keep other service instances running.

Before maintenance window

• First you should execute required DB schema migrations. To achieve that without interrupting your existing installation, install new Cinder code in new venv or a container and run the DB sync (cinder-manage db sync). These schema change operations should have minimal or no effect on performance, and should not cause any operations to fail.

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

During maintenance window

1. The first service is cinder-scheduler. It is load-balanced by the message queue, so the only thing you need to worry about is to shut it down gracefully (using SIGTERM signal) to make sure it will finish all the requests being processed before shutting down. Then you should upgrade the code and restart the service.

2. Repeat first step for all of your cinder-scheduler services.

3. Then you proceed to upgrade cinder-volume services. The problem here is that due to Active/Passive character of this service, you’re unable to run multiple instances of cinder-volume managing a single volume backend. This means that there will be a moment when you won’t have any cinder-volume in your deployment and you want that disruption to be as short as possible.

   Note

   The downtime here is non-disruptive as long as it doesn’t exceed the service heartbeat timeout. If you don’t exceed that, then cinder-schedulers will not notice that cinder-volume is gone and the message queue will take care of queuing any RPC messages until cinder-volume is back.

   To make sure it’s achieved, you can either lengthen the timeout by tweaking service_down_time value in cinder.conf, or prepare upgraded cinder-volume on another node and do a very quick switch by shutting down older service and starting the new one just after that.

   Also note that in case of A/P HA configuration you need to make sure both primary and secondary c-vol have the same hostname set (you can override it using host option in cinder.conf), so both will be listening on the same message queue and will accept the same messages.

4. Repeat third step for all cinder-volume services.

5. Now we should proceed with (optional) cinder-backup services. You should upgrade them in the same manner like cinder-scheduler.

   Note

   Backup operations are time consuming, so shutting down a c-bak service without interrupting ongoing requests can take time. It may be useful to disable the service first using cinder service-disable command, so it won’t accept new requests, and wait a reasonable amount of time until all the in-progress jobs are completed. Then you can proceed with the upgrade. To make sure the backup service finished all the ongoing requests, you can check the service logs.

   Note

   Until Liberty cinder-backup was tightly coupled with cinder-volume service and needed to coexist on the same physical node. This is not true starting with Mitaka version. If you’re still keeping that coupling, then your upgrade strategy for cinder-backup should be more similar to how cinder-volume is upgraded.

6. cinder-api services should go last. In HA deployment you’re typically running them behind a load balancer (e.g. HAProxy), so you need to take one service instance out of the balancer, shut it down, upgrade the code and dependencies, and start the service again. Then you can plug it back into the load balancer.

   Note

   You may want to start another instance of older c-api to handle the load while you’re upgrading your original services.

7. Then you should repeat step 6 for all of the cinder-api services.

After maintenance window

• Once all services are running the new code, double check in the DB that there are no old orphaned records in services table (Cinder doesn’t remove the records when service is gone or service hostname is changed, so you need to take care of that manually; you should be able to distinguish dead records by looking at when the record was updated). Cinder is basing its RPC version detection on that, so stale records can prevent you from going forward.

• Now all services are upgraded, we need to send the SIGHUP signal, so all the services clear any cached service version data. When a new service starts, it automatically detects which version of the service’s RPC protocol to use, and will downgrade any communication to that version. Be advised that cinder-api service doesn’t handle SIGHUP so it needs to be restarted. It’s best to restart your cinder-api services as last ones, as that way you make sure API will fail fast when user requests new features on a deployment that’s not fully upgraded (new features can fail when RPC messages are backported to lowest common denominator). Order of the rest of the services shouldn’t matter.

• Now 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.

• At this point, you must also ensure you update the configuration, to stop using any deprecated features or options, and perform any required work to transition to alternative features. All the deprecated options should be supported for one cycle, but should be removed before your next upgrade is performed.

• Since Ocata, you also need to run cinder-manage db online_data_migrations command to make sure data migrations are applied. The tool lets you limit the impact of the data migrations by using --max_count option to limit number of migrations executed in one run. If this option is used, the exit status will be 1 if any migrations were successful (even if others generated errors, which could be due to dependencies between migrations). The command should be rerun while the exit status is 1. If no further migrations are possible, the exit status will be 2 if some migrations are still generating errors, which requires intervention to resolve. The command should be considered completed successfully only when the exit status is 0. You need to complete all of the migrations before starting upgrade to the next version (e.g. you need to complete Ocata’s data migrations before proceeding with upgrade to Pike; you won’t be able to execute Pike’s DB schema migrations before completing Ocata’s data migrations).