When upgrading or downgrading OpenStack, it is possible to use package based management or image-based management. Containerizing OpenStack is meant to optimize image-based management of OpenStack. Containerizing OpenStack solves a manageability and availability problem with the current state of the art deployment systems in OpenStack.
Current state of the art deployment systems use either image based or package based upgrade.
Image based upgrades are utilized by TripleO. When TripleO updates a system, it creates an image of the entire disk and deploys that rather than just the parts that compose the OpenStack deployment. This results in significant loss of availability. Further running VMs are shut down in the imaging process. However, image based systems offer atomicity, because all related software for a service is updated in one atomic action by reimaging the system.
Other systems use package based upgrade. Package based upgrades suffer from a non-atomic nature. An update may update 1 or more RPM packages. The update process could fail for any number of reasons, and there is no way to back out the existing changes. Typically in an OpenStack deployment it is desirable to update a service that does one thing including it's dependencies as an atomic unit. Package based upgrades do not offer atomicity.
To solve this problem, containers can be used to provide an image-based update approach which offers atomic upgrade of a running system with minimal interruption in service. A rough prototype of compute upgrade [1] shows approximately a 10 second window of unavailability during a software update. The prototype keeps virtual machines running without interruption.
- Upgrade or rollback OpenStack deployments atomically. End-user wants to change the running software versions in her system to deploy a new upstream release without interrupting service for significant periods.
- Upgrade OpenStack based by component. End-user wants to upgrade her system in fine-grained chunks to limit damage from a failed upgrade.
- Rollback OpenStack based by component. End-user experienced a failed upgrade and wishes to rollback to the last known good working version.
An OpenStack deployment based on containers are represented in a tree structure with each node representing a container set, and each leaf representing a container.
The full properties of a container set:
- A container set is composed of one or more container subsets or one or more individual containers
- A container set provides a single logical service
- A container set is managed as a unit during startup, shutdown, and version
- Each container set is launched together as one unit
- A container set with subsets is launched as one unit including all subsets
- A container set is not atomically managed
- A container set provides appropriate hooks for high availability monitoring
The full properties of a container:
- A container is atomically upgraded or rolled back
- A container includes a monotonically increasing generation number to identify the container's age in comparison with other containers
- A container has a single responsibility
- A container may be super-privileged when it needs significant access to the
host including:
- the network namespace of the host
- The UUID namespace of the host
- The IPC namespace of the host
- Filesystem sharing of the host for persistent storage
- A container may lack any privileges when it does not require significant access to the host.
- A container should include a check function for evaluating its own health.
- A container will include proper PID 1 handling for reaping exited child processes.
The top level container sets are composed of:
- database control
- messaging control
- high availability control
- OpenStack interface
- OpenStack control
- OpenStack compute operation
- OpenStack network operation
- OpenStack storage operation
The various container sets are composed in more detail as follows:
- Database control
- galera
- mariadb
- mongodb
- Messaging control
- rabbitmq
- High availability control
- HAProxy
- keepalived
- OpenStack interface
- keystone
- glance-api
- nova-api
- ceilometer-api
- heat-api
- OpenStack control
- glance-controller
- glance-registry
- nova-controller
- nova-conductor
- nova-scheduler
- metadata-service
- cinder-controller
- neutron-controller
- neutron-server
- ceilometer-controller
- ceilometer-alarm
- ceilometer-base
- ceilometer-central
- ceilometer-collector
- ceilometer-notification
- heat-controller
- heat-engine
- glance-controller
- OpenStack compute operation
- nova-compute
- nova-libvirt
- neutron-agents-linux-bridge
- neutron-agents-ovs
- OpenStack network operation
- dhcp-agent
- l3-agent
- metadata-agent
- lbaas-agent
- fwaas-agent
- OpenStack storage operation
- Cinder
- Swift
- swift-account
- swift-base
- swift-container
- swift-object
- swift-proxy-server
In order to achieve the desired results, we plan to permit super-privileged containers. A super-privileged container is defined as any container launched with the --privileged=true flag to docker that:
- bind-mounts specific security-crucial host operating system directories with -v. This includes nearly all directories in the filesystem except for leaf directories with no other host operating system use.
- shares any namespace with the --ipc=host, --pid=host, or --net=host flags
We will not use the Docker EXPOSE operation since all containers will use --net=host. One motive for using --net=host is it is inherently simpler. A different motive for not using EXPOSE is the 20 microsecond penalty applied to every packet forwarded and returned by docker-proxy. If EXPOSE functionality is desired, it can be added back by referencing the default list of OpenStack ports to each Dockerfile: Firewalls and default ports.
We will use the docker flag --restart=always to provide some measure of high availability for the individual containers and ensure they operate correctly as currently designed.
A host tool will run and monitor the container's built-in check script via docker exec to validate the container is operational on a pre-configured timer. If the container does not pass its healthcheck operation, it should be restarted.
Integration of metadata with fig or a similar single node Docker orchestration tool will be implemented. Even though fig executes on a single node, the containers will be designed to run multi-node and the deploy tool should take some form of information to allow it to operate multi-node. The deploy tool should take a set of key/value pairs as inputs and convert them into inputs into the environment passed to Docker. These key/value pairs could be a file or environment variables. We will not offer integration with multi-node scheduling or orchestration tools, but instead expect our consumers to manage each bare metal machine using our fig or similar in nature tool integration.
Any contributions from the community of the required metadata to run these containers using a multi-node orchestration tool will be warmly received but generally won't be maintained by the core team.
The technique for launching the deploy script is not handled by Kolla. This is a problem for a higher level deployment tool such as TripleO or Fuel to tackle.
Logs from the individual containers will be retrievable in some consistent way.
Container usage with super-privileged mode may possibly impact security. For example, when using --net=host mode and bind-mounting /run which is necessary for a compute node, it is possible that a compute breakout could corrupt the host operating system.
To mitigate security concerns, solutions such as SELinux and AppArmor should be used where appropriate to contain the security privileges of the containers.
The upgrade or downgrade process changes from a multi-hour outage to a 10 second outage across the system.
Primary assignee:
kolla maintainers
- Container Sets
- Containers
- A minimal proof of concept single-node fig deployment integration
- A minimal proof of concept fig healthchecking integration
Functional tests will be implemented in the OpenStack check/gating system to automatically check that containers pass each container's functional tests stored in the project's repositories.
The documentation impact is unclear as this project is a proof of concept with no clear delivery consumer.