moby--moby/libnetwork/docs/design.md

Design

The vision and goals of libnetwork are highlighted in roadmap. This document describes how libnetwork has been designed in order to achieve this. Requirements for individual releases can be found on the Project Page.

Many of the design decisions are inspired by the learnings from the Docker networking design as of Docker v1.6. Please refer to this Docker v1.6 Design document for more information on networking design as of Docker v1.6.

Goal

libnetwork project will follow Docker and Linux philosophy of developing small, highly modular and composable tools that works well independently. Libnetwork aims to satisfy that composable need for Networking in Containers.

The Container Network Model

Libnetwork implements Container Network Model (CNM) which formalizes the steps required to provide networking for containers while providing an abstraction that can be used to support multiple network drivers. The CNM is built on 3 main components (shown below)

Sandbox

A Sandbox contains the configuration of a container's network stack. This includes management of the container's interfaces, routing table and DNS settings. An implementation of a Sandbox could be a Linux Network Namespace, a FreeBSD Jail or other similar concept. A Sandbox may contain many endpoints from multiple networks.

Endpoint

An Endpoint joins a Sandbox to a Network. An implementation of an Endpoint could be a veth pair, an Open vSwitch internal port or similar. An Endpoint can belong to only one network but may only belong to one Sandbox.

Network

A Network is a group of Endpoints that are able to communicate with each-other directly. An implementation of a Network could be a Linux bridge, a VLAN, etc. Networks consist of many endpoints.

CNM Objects

NetworkController NetworkController object provides the entry-point into libnetwork that exposes simple APIs for the users (such as Docker Engine) to allocate and manage Networks. libnetwork supports multiple active drivers (both inbuilt and remote). NetworkController allows user to bind a particular driver to a given network.

Driver Driver is not an user visible object, but drivers provides the actual implementation that makes network work. NetworkController however provides an API to configure any specific driver with driver-specific options/labels that is transparent to libnetwork, but can be handled by the drivers directly. Drivers can be both inbuilt (such as Bridge, Host, None & overlay) and remote (from plugin providers) to satisfy various usecases & deployment scenarios. At this point, the Driver owns a network and is responsible for managing the network (including IPAM, etc.). This can be improved in the future by having multiple drivers participating in handling various network management functionalities.

Network Network object is an implementation of the CNM : Network as defined above. NetworkController provides APIs to create and manage Network object. Whenever a Network is created or updated, the corresponding Driver will be notified of the event. LibNetwork treats Network object at an abstract level to provide connectivity between a group of end-points that belong to the same network and isolate from the rest. The Driver performs the actual work of providing the required connectivity and isolation. The connectivity can be within the same host or across multiple-hosts. Hence Network has a global scope within a cluster.

Endpoint Endpoint represents a Service Endpoint. It provides the connectivity for services exposed by a container in a network with other services provided by other containers in the network. Network object provides APIs to create and manage endpoint. An endpoint can be attached to only one network. Endpoint creation calls are made to the corresponding Driver which is responsible for allocating resources for the corresponding Sandbox. Since Endpoint represents a Service and not necessarily a particular container, Endpoint has a global scope within a cluster as well.

Sandbox Sandbox object represents container's network configuration such as ip-address, mac-address, routes, DNS entries. A Sandbox object is created when the user requests to create an endpoint on a network. The Driver that handles the Network is responsible to allocate the required network resources (such as ip-address) and pass the info called SandboxInfo back to libnetwork. libnetwork will make use of OS specific constructs (example: netns for Linux) to populate the network configuration into the containers that is represented by the Sandbox. A Sandbox can have multiple endpoints attached to different networks. Since Sandbox is associated with a particular container in a given host, it has a local scope that represents the Host that the Container belong to.

CNM Attributes

Options Options provides a generic and flexible mechanism to pass Driver specific configuration option from the user to the Driver directly. Options are just key-value pairs of data with key represented by a string and value represented by a generic object (such as golang interface{}). Libnetwork will operate on the Options ONLY if the key matches any of the well-known Label defined in the net-labels package. Options also encompasses Labels as explained below. Options are generally NOT end-user visible (in UI), while Labels are.

Labels Labels are very similar to Options & in fact they are just a subset of Options. Labels are typically end-user visible and are represented in the UI explicitly using the --labels option. They are passed from the UI to the Driver so that Driver can make use of it and perform any Driver specific operation (such as a subnet to allocate IP-Addresses from in a Network).

CNM Lifecycle

Consumers of the CNM, like Docker for example, interact through the CNM Objects and its APIs to network the containers that they manage.

0. Drivers registers with NetworkController. Build-in drivers registers inside of LibNetwork, while remote Drivers registers with LibNetwork via Plugin mechanism. (plugin-mechanism is WIP). Each driver handles a particular networkType.

1. NetworkController object is created using libnetwork.New() API to manage the allocation of Networks and optionally configure a Driver with driver specific Options.

2. Network is created using the controller's NewNetwork() API by providing a name and networkType. networkType parameter helps to choose a corresponding Driver and binds the created Network to that Driver. From this point, any operation on Network will be handled by that Driver.

3. controller.NewNetwork() API also takes in optional options parameter which carries Driver-specific options and Labels, which the Drivers can make use for its purpose.

4. network.CreateEndpoint() can be called to create a new Endpoint in a given network. This API also accepts optional options parameter which drivers can make use of. These 'options' carry both well-known labels and driver-specific labels. Drivers will in turn be called with driver.CreateEndpoint and it can choose to reserve IPv4/IPv6 addresses when an Endpoint is created in a Network. The Driver will assign these addresses using InterfaceInfo interface defined in the driverapi. The IP/IPv6 are needed to complete the endpoint as service definition along with the ports the endpoint exposes since essentially a service endpoint is nothing but a network address and the port number that the application container is listening on.

5. endpoint.Join() can be used to attach a container to an Endpoint. The Join operation will create a Sandbox if it doesn't exist already for that container. The Drivers can make use of the Sandbox Key to identify multiple endpoints attached to a same container. This API also accepts optional options parameter which drivers can make use of.

• Though it is not a direct design issue of LibNetwork, it is highly encouraged to have users like Docker to call the endpoint.Join() during Container's Start() lifecycle that is invoked before the container is made operational. As part of Docker integration, this will be taken care of.
• One of a FAQ on endpoint join() API is that, why do we need an API to create an Endpoint and another to join the endpoint.
  • The answer is based on the fact that Endpoint represents a Service which may or may not be backed by a Container. When an Endpoint is created, it will have its resources reserved so that any container can get attached to the endpoint later and get a consistent networking behaviour.

6. endpoint.Leave() can be invoked when a container is stopped. The Driver can cleanup the states that it allocated during the Join() call. LibNetwork will delete the Sandbox when the last referencing endpoint leaves the network. But LibNetwork keeps hold of the IP addresses as long as the endpoint is still present and will be reused when the container(or any container) joins again. This ensures that the container's resources are reused when they are Stopped and Started again.

7. endpoint.Delete() is used to delete an endpoint from a network. This results in deleting an endpoint and cleaning up the cached sandbox.Info.

8. network.Delete() is used to delete a network. LibNetwork will not allow the delete to proceed if there are any existing endpoints attached to the Network.

Implementation Details

Networks & Endpoints

LibNetwork's Network and Endpoint APIs are primarily for managing the corresponding Objects and book-keeping them to provide a level of abstraction as required by the CNM. It delegates the actual implementation to the drivers which realize the functionality as promised in the CNM. For more information on these details, please see the drivers section

Sandbox

Libnetwork provides a framework to implement of a Sandbox in multiple operating systems. Currently we have implemented Sandbox for Linux using namespace_linux.go and configure_linux.go in sandbox package This creates a Network Namespace for each sandbox which is uniquely identified by a path on the host filesystem. Netlink calls are used to move interfaces from the global namespace to the Sandbox namespace. Netlink is also used to manage the routing table in the namespace.

Drivers

API

Drivers are essentially an extension of libnetwork and provides the actual implementation for all of the LibNetwork APIs defined above. Hence there is an 1-1 correspondence for all the Network and Endpoint APIs, which includes :

• driver.Config
• driver.CreateNetwork
• driver.DeleteNetwork
• driver.CreateEndpoint
• driver.DeleteEndpoint
• driver.Join
• driver.Leave

These Driver facing APIs makes use of unique identifiers (networkid,endpointid,...) instead of names (as seen in user-facing APIs).

The APIs are still work in progress and there can be changes to these based on the driver requirements especially when it comes to Multi-host networking.

Driver semantics

• Driver.CreateEndpoint

This method is passed an interface EndpointInfo, with methods Interface and AddInterface.

If the value returned by Interface is non-nil, the driver is expected to make use of the interface information therein (e.g., treating the address or addresses as statically supplied), and must return an error if it cannot. If the value is nil, the driver should allocate exactly one fresh interface, and use AddInterface to record them; or return an error if it cannot.

It is forbidden to use AddInterface if Interface is non-nil.

Implementations

Libnetwork includes the following driver packages:

• null
• bridge
• overlay
• remote

Null

The null driver is a noop implementation of the driver API, used only in cases where no networking is desired. This is to provide backward compatibility to the Docker's --net=none option.

Bridge

The bridge driver provides a Linux-specific bridging implementation based on the Linux Bridge. For more details, please see the Bridge Driver documentation.

Overlay

The overlay driver implements networking that can span multiple hosts using overlay network encapsulations such as VXLAN. For more details on its design, please see the Overlay Driver Design.

Remote

The remote package does not provide a driver, but provides a means of supporting drivers over a remote transport. This allows a driver to be written in a language of your choice. For further details, please see the Remote Driver Design.