zun/specs/kuryr-integration.rst
Kien Nguyen 76057450b5 Fix docs & specs erros.
- Fix erros when run tox -edocs (except one spec - cinder-integration.rst,
  will be fixed when merge patch set [1])
- Update urls.

[1] https://review.openstack.org/#/c/468658/

Change-Id: I05865708ef356f17a388eace234ac004dd1a364f
2017-07-15 16:10:06 +07:00

7.8 KiB

Kuryr Integration

Related Launchpad Blueprint:

https://blueprints.launchpad.net/zun/+spec/kuryr-integration

Zun provides APIs for managing application containers, and the implementation of the APIs is provided by drivers. Currently, Zun has two built-in drivers: the native Docker driver and the Nova Docker driver. The Nova driver leverages existing Nova capability to provide networking for containers. However, the native Docker driver doesn't have an ideal solution for networking yet.

This spec proposed to leverage Kuryr-libnetwork [1] for providing networking for containers. Generally speaking, Kuryr-libnetwork is a remote Docker networking plugin that receives requests from Docker engine and interfaces with Neutron for managing the network resources. Kuryr-libnetwork provides several features, such as IPAM, Neutron port binding, etc., all of which could be leveraged by Zun.

Problem description

Currently, the native Docker driver doesn't integrate with Neutron. It uses the default networking capability provided by Docker engine. Containers created by that driver has limited networking capabilities, and they are not able to communicate with other OpenStack resources (i.e. Nova instances).

Proposed change

  1. Introduce a network abstraction in Zun. Implement an API for users to create/delete/manage container networks backed by Kuryr. If the container runtime is Docker, creating a container network will call docker network APIs to create a Docker network by using the pre-created Neutron resources (this capability is provided by Kuryr [2][3]). Deleting a container network will be similar as create.
  2. Support creating a container from a network. If a user specifies a network on creating a container, the container will be created from the network. If not, the driver will take care the networking of the container. For example, some drivers might choose to create the container from a default network that might be specified in a config file or hard-coded. It is up to individual driver to decide how to create a container from a network. On the Zun's Docker driver, this is implemented by adding --net=<ID> option when creating the sandbox [4] of the container.

The typical workflow will be as following:

  1. Users call Zun APIs to create a container network by passing a name/uuid of a neutron network:

    $ zun network-create --neutron-net private --name foo
  2. After receiving this request, Zun will make several API calls to Neutron to retrieve the necessary information about the specified network ('private' in this example). In particular, Zun will list all the subnets that belong to 'private' network. The number of subnets under a network should only be one or two. If the number of subnets is two, they must be a ipv4 subnet and a ipv6 subnet respectively. Zun will retrieve the cidr/gateway/subnetpool of each subnet and pass these information to Docker to create a Docker network. The API call will be similar to:

    $ docker network create -d kuryr --ipam-driver=kuryr \
                      --subnet <ipv4_cidr> \
                      --gateway <ipv4_gateway> \
                      -ipv6 --subnet <ipv6_cidr> \
                      --gateway <ipv6_gateway> \
                      -o neutron.net.uuid=<network_uuid> \
                      -o neutron.pool.uuid=<ipv4_pool_uuid> \
                      --ipam-opt neutron.pool.uuid=<ipv4_pool_uuid> \
                      -o neutron.pool.v6.uuid=<ipv6_pool_uuid> \
                      --ipam-opt neutron.pool.v6.uuid=<ipv6_pool_uuid> \
                      foo

NOTE: In this step, docker engine will check the list of registered network plugin and find the API endpoint of Kuryr, then make a call to Kuryr to create a network with existing Neutron resources (i.e. network, subnetpool, etc.). This example assumed that the Neutron resources were pre-created by cloud administrator (which should be the case at most of the clouds). If this is not true, users need to manually create the resources.

  1. Users call Zun APIs to create a container from the container network 'foo':

    $ zun run --net=foo nginx
  2. Under the hood, Zun will perform several steps to configure the networking. First, call neutron API to create a port from the specified neutron network:

    $ neutron port-create private
  3. Then, Zun will retrieve information of the created neutron port and retrieve its IP address(es). A port could have one or two IP addresses: a ipv4 address and/or a ipv6 address. Then, call Docker APIs to create the container by using the IP address(es) of the neutron port. This is equivalent to:

    $ docker run --net=foo kubernetes/pause --ip <ipv4_address> \
                                            --ip6 <ipv6_address>

NOTE: In this step, docker engine will make a call to Kuryr to setup the networking of the container. After receiving the request from Docker, Kuryr will perform the necessary steps to connect the container to the neutron port. This might include something like: create a veth pair, connect one end of the veth pair to the container, connect the other end of the veth pair a neutron-created bridge, etc.

  1. Users calls Zun API to list/show the created network(s):

    $ zun network-list
    $ zun network-show foo
  2. Upon completion, users calls Zun API to remove the container and network:

    $ zun delete <container_id>
    $ zun network-delete foo

Alternatives

  1. Directly integrate with Neutron (without Kuryr-libnetwork). This approach basically re-invented Kuryr functionalities in Zun, which is unnecessary.
  2. Use alternative networking solution (i.e. Flannel) instead of Neutron. This doesn't provide a good OpenStack integration.

Data model impact

  • Create a 'network' table. Each entry in this table is a record of a network. A network must belong to an OpenStack project so there will be a 'project_id' column in this table.
  • Create a 'network_attachment' table. Each entry in this table is a record of an attachment between a network and a container. In fact, this table defines a many-to-many relationship between networks and containers.

REST API impact

  1. Add a new API endpoint /networks to the REST API interface.
  2. In the API endpoint of creating a container, add a new option to specify the network where the container will be created from.

Security impact

None

Notifications impact

None

Other end user impact

None

Performance Impact

None

Other deployer impact

Deployers need to deploy a Kuryr-libnetwork as a prerequisites of using this feature.

Developer impact

None

Implementation

Assignee(s)

Primary assignee: Hongbin Lu

Other contributors: Sudipta Biswas

Work Items

  1. Implement a new API endpoint for networks.
  2. Extend the Docker driver to support creating containers from a network.
  3. Implement unit/integration test.
  4. Document the new network API.

Dependencies

Add a dependency to Kuryr-libnetwork and Neutron

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://git.openstack.org/cgit/openstack/kuryr-libnetwork

[2] https://blueprints.launchpad.net/kuryr/+spec/existing-neutron-network

[3] https://blueprints.launchpad.net/kuryr-libnetwork/+spec/existing-subnetpool

[4] https://git.openstack.org/cgit/openstack/zun/tree/specs/container-sandbox.rst