zun/specs/container-composition.rst

Container Composition

Related Launchpad Blueprint:

https://blueprints.launchpad.net/zun/+spec/introduce-compose

Kubernetes Pod[1] or Docker compose[2] are popular for deploying applications or application components that span multiple containers. It is a basic unit for scheduler, resource allocation. This spec proposes to support a similar feature in Zun, we can put multiple containers, a sandbox and other related resources into one unit, we name the unit capsule.

The containers in a capsule are relatively tightly coupled, they share the capsule's context like Linux namespaces, cgroups and etc, and they work together closely to form a cohesive unit of service.

Problem description

Currently running or deploying one container to do the operation is not a very effective way in micro services, while multiple different containers run as an integration has widely used in different scenarios, such as pod in Kubernetes. The pod has the independent network, storage, while the compose has an easy way to defining and running multi-container Docker applications. They are becoming the basic unit for container application scenarios.

Nowadays Zun doesn't support creating and running multiple containers as an integration. So we will introduce the new Object capsule to realize this function. capsule is the basic unit for zun to support service to external.

The capsule will be designed based on some similar concepts such as pod and compose. For example, capsule can be specified in a yaml file that might be similar to the format of k8s pod manifest. However, the specification of capsule will be exclusive to Zun. The details will be showed in the following section.

Proposed change

A capsule has the following properties: * Structure: It can contains one or multiple containers, and has a sandbox container which will support the network namespace for the capsule. * Scheduler: Containers inside a capsule are scheduled as a unit, thus all containers inside a capsule is co-located. All containers inside a capsule will be launched in one compute host. * Network: Containers inside a capsule share the same network namespace, so they share IP address(es) and can find each other via localhost by using different remapping network port. Capsule IP address(es) will re-use the sandbox IP. Containers communication between different capsules will use capsules IP and port. * LifeCycle: Capsule has different status: Starting: Capsule is created, but one or more container inside the capsule is being created. Running: Capsule is created, and all the containers are running. Finished: All containers inside the capsule have successfully executed and exited. Failed: Capsule creation is failed * Restart Policy: Capsule will have a restart policy just like container. The restart policy relies on container restart policy to execute. * Health checker: In the first step of realization, container inside the capsule will send its status to capsule when its status changed. * Upgrade and rollback: Upgrade: Support capsule update(different from zun update). That means the container image will update, launch the new capsule from new image, then destroy the old capsule. The capsule IP address will change. For Volume, need to clarify it after Cinder integration. Rollback: When update failed, rollback to it origin status. * CPU and memory resources: Given that host resource allocation, cpu and memory support will be implemented.

Implementation:

1. Introduce a new abstraction called capsule. It represents a tightly coupled unit of multiple containers and other resources like sandbox. The containers in a capsule shares the capsule's context like Linux namespaces and cgroups.
2. Support the CRUD operations against capsule object, capsule should be a basic unit for scheduling and spawning. To be more specific, all containers in a capsule should be scheduled to and spawned on the same host. Server side will keep the information in DB.
3. Add functions about yaml file parser in the CLI side. After parsing the yaml, send the REST to API server side, scheduler will decide which host to run the capsule.
4. Introduce new REST API for capsule. The capsule creation workflow is: CLI Parsing capsule information from yaml file --> API server do the CRUD operation, call scheduler to launch the capsule, from Cinder to get volume, from Kuryr to get network support--> Compute host launch the capsule, attach the volume--> Send the status to API server, update the DB.
5. Capsule creation will finally depend on the backend container driver. Now choose Docker driver first.
6. Define a yaml file structure for capsule. The yaml file will be compatible with Kubernetes pod yaml file, at the same time Zun will define the available properties, metadata and template of the yaml file. In the first step, only essential properties will be defined.

The diagram below offers an overview of the architecture of capsule:

::

Sandbox +-------------+ +-------------+ +-------------+ | | | | | | | Container | | Container | | Container | | | | | | | +-------------+ +-------------+ +-------------+ +----------+ +----------+ | | | | | Volume | | Volume | | | | | +----------+ +----------+

Yaml format for capsule:

Sample capsule: .. code-block:: yaml apiVersion: beta kind: capsule metadata: name: capsule-example lables: app: web restartPolicy: Always hostSelector: node1 spec: containers: - image: ubuntu:trusty command: ["echo"] args: ["Hello World"] imagePullPolicy: Always imageDriver: Glance workDir: /root labels: app: web volumeMounts: - name: volume1 mountPath: /root/mnt readOnly: True ports: - name: nginx-port containerPort: 80 hostPort: 80 protocol: TCP env: PATH: /usr/local/bin resources: allocation: cpu: 1 memory: 2GB volumes: - name: volume1 drivers: cinder driverOptions: options size: 5GB volumeType: type1 image: ubuntu-xenial

Capsule fields: * apiVersion(string): the first version is beta * kind(string): the flag to show yaml file property * metadata(ObjectMeta): metadata Object * spec(CapsuleSpec): capsule specifications * restartPolicy(string): [Always | Never | OnFailure], by default is Always * hostSelector(string): Specify the host that will launch the capsule

ObjectMeta fields: * name(string): capsule name * lables(dict, name: string): labels for capsule

CapsuleSpec fields: * containers(Containers array): containers info array, one capsule have multiple containers * volumes(Volumes array): volume information

Containers fields: * name(string): name for container * image(string): container image for container * imagePullPolicy(string): [Always | Never | IfNotPresent] * imageDriver(string): glance or dockerRegistory, by default is according to zun configuration * command(string): container command when starting * args(string): container args for the command * workDir(string): workDir for the container * labels(dict, name:string): labels for the container * volumeMounts(VolumnMounts array): volumeMounts information for container * ports(Ports array): Port mapping information for container * env(dict, name:string): environment variables for container * resources(RecourcesObject): resources that container needed

VolumnMounts fields: * name(string): volume name that listed in below field "volumes" * mountPath(string): mount path that in the container, absolute path * readOnly(boolean): read only flags

Ports fields: * name(string): port name, optional * containerPort(int): port number that container need to listen * hostPort(int): port number that capsule need to listen * protocol(string): TCP or UDP, by default is TCP

RecourcesObject fields: * allocation(AllocationObject): the resources that the capsule needed

AllocationObject: * cpu(string): cpu resources, cores number * memory(string): memory resources, MB or GB

Volumes fields: * name(string): volume name * driver(string): volume drivers * driverOptions(string): options for volume driver * size(string): volume size * volumeType(string): volume type that cinder need. by default is from cinder config * image(string): cinder needed to boot from image

Alternatives

1. Abstract all the information from yaml file and implement the capsule CRUD in client side.
2. Implement the CRUD in server side.

Data model impact

• Add a field to container to store the id of the capsule which include the container
• Create a 'capsule' table. Each entry in this table is a record of a capsule.

Introduce the capsule Object:
    fields = {
    'capsuleVersion': fields.StringField(nullable=True),
    'kind': fields.StringField(nullable=True),
    'id': fields.IntegerField(),
    'uuid': fields.UUIDField(nullable=True),
    'name': fields.StringField(nullable=True),
    'project_id': fields.StringField(nullable=True),
    'user_id': fields.StringField(nullable=True),

    'status': z_fields.ContainerStatusField(nullable=True),
    'status_reason': fields.StringField(nullable=True),

    # conclude the readable message that show why capsule is in this status
    # 'key': 'value'--> 'time':'message'
    'message': fields.DictOfStringsField(nullable=True),
    'startTime': fields.StringField(nullable=True),

    'cpu': fields.FloatField(nullable=True),
    'memory': fields.StringField(nullable=True),
    'task_state': z_fields.TaskStateField(nullable=True),
    'host': fields.StringField(nullable=True),
    'restart_policy': fields.DictOfStringsField(nullable=True),

    'meta': fields.DictOfStringsField(nullable=True),
    'volumes': fields.DictOfStringsField(nullable=True),
    'ip': fields.StringField(nullable=True),
    'labels': fields.DictOfStringsField(nullable=True),
    'ports': z_fields.ListOfIntegersField(nullable=True),
    'hostname': fields.StringField(nullable=True),
}

REST API impact

• Add a new API endpoint /capsule to the REST API interface.

* Capsule API: Capsule consider to support multiple operations as container composition. * Container API: Many container API will be extended to capsule. Here in this section will define the API usage range.

Capsule API: list <List all the capsule, add parameters about list capsules with the same labels> create <-f yaml file><-f directory> describe <display the details state of one or more resource> delete <capsule name> <-l name=label-name> <–all> run <--capsule ... container-image> If "--capsule .." is set, the container will be created inside the capsule. Otherwise, it will be created as normal.

Container API: * show/list allow all containers * create/delete allow bare container only (disallow in-capsule containers) * attach/cp/logs/top allow all containers * start/stop/restart/kill/pause/unpause allow bare container only (disallow in-capsule containers) * update for container in the capsule, need <--capsule> params. Bare container doesn't need.

Security impact

None

Notifications impact

Need to support "zun notification" for capsule events

Other end user impact

None

Performance Impact

None

Other deployer impact

None

Developer impact

None

Implementation

The implementation is divided into the following parts: 1. Define the capsule data structure. Take Kubernetes Pod as a reference. 2. Define the yaml structure for capsule, add the parser for the yaml file. The parser realization is in CLI. CLI parse info from yaml and then send to API server. 3. Implement a new API endpoint for capsule, including capsule life cycle and information. 4. Implement the API server side, including DB CRUD, compute node scheduler, etc. 5. Implement the compute server side, now using Docker Driver first. The first step will just realize the several containers in the same sandbox which have the same network namespace. The storage share in the capsule will be added after Cinder integration.

We will split the implementation into several blueprints for easy task tracking.

Assignee(s)

Primary assignee: Wenzhi Yu <yuywz> Kevin Zhao <kevinz>

Work Items

1. Implement a new API endpoint for capsules.
2. Implement unit/integration test.
3. Document the new capsule API.

Dependencies

1. Need to add support for select host to launch capsule
2. Need to add support for port mapping
3. Need to support "zun notification" for capsule events

Testing

Each patch will have unit tests, and Tempest functional tests covered.

Documentation Impact

A set of documentation for this new feature will be required.

References

[1] https://kubernetes.io/ [2] https://docs.docker.com/compose/ [3] https://etherpad.openstack.org/p/zun-container-composition