openstack-helm/doc/source/philosophy.rst

Philosophy

Resiliency Philosophy

One of the goals of this project is to produce a set of charts that can be used in a production setting to deploy and upgrade OpenStack. To achieve this goal, all components must be resilient, including both OpenStack and Infrastructure components leveraged by this project. In addition, this also includes Kubernetes itself. It is part of our mission to ensure that all infrastructure components are highly available and that a deployment can withstand a physical host failure out of the box. This means that:

• OpenStack components need to support and deploy with multiple replicas out of the box to ensure that each chart is deployed as a single-unit production ready first class citizen (unless development mode is enabled).
• Infrastructure elements such as Ceph, RabbitMQ, Galera (MariaDB), Memcached, and all others need to support resiliency and leverage multiple replicas for resiliency where applicable. These components also need to validate that their application level configurations (for instance the underlying Galera cluster) can tolerate host crashes and withstand physical host failures.
• Scheduling annotations need to be employed to ensure maximum resiliency for multi-host environments. They also need to be flexible to allow all-in-one deployments. To this end, we promote the usage of podAntiAffinity.preferredDuringSchedulingIgnoredDuringExecution for most infrastructure elements.
• We make the assumption that we can depend on a reliable implementation of centralized storage to create PVCs within Kubernetes to support resiliency and complex application design. Today, this is provided by the included Ceph chart. There is much work to do when making even a single backend production ready. We have chosen to focus on bringing Ceph into a production ready state, which includes handling real world deployment scenarios, resiliency, and pool configurations. In the future we would like to support more options for hardened backend PVC's. In the future, we would like to offer flexibility in choosing a hardened backend.
• We will document the best practices for running a resilient Kubernetes cluster in production. This includes documenting the steps necessary to make all components resilient, such as Etcd and SkyDNS where possible, and point out gaps due to missing features.

Scaling Philosophy

Scaling is another first class citizen in openstack-helm. We will be working to ensure that we support various deployment models that can support hyperscale, such as:

• Ensuring that by default, clusters include multiple replicas to verify that scaling issues are identified early and often (unless development mode is enabled).
• Ensuring that every chart can support more then one replica and allowing operators to override those replica counts. For some applications, this means that they support clustering.
• Ensuring clustering style applications are not limited to fixed replica counts. For instance, we want to ensure that we can support n Galera members and have those scale linearly, within reason, as opposed to only supporting a fixed count.
• Duplicate charts of the same type within the same namespace. For example, deploying rabbitmq twice, to the openstack namespace resulting in two fully functioning clusters.
• Allowing charts to be deployed to a diverse set of namespaces. For example, allowing infrastructure to be deployed in one namespace and OpenStack in another, or deploying each chart in its own namespace.
• Supporting hyperscale configurations that call for per-component infrastructure, such as a dedicated database and RabbitMQ solely for Ceilometer, or even dedicated infrastructure(s) for every component you deploy. It is unique, large scale deployment designs such as this that only become practical under a Kubernetes/Container framework and we want to ensure that we can support them.