| Network Working Group | A. Clemm |
| Internet-Draft | Futurewei |
| Intended status: Informational | L. Ciavaglia |
| Expires: May 7, 2020 | Nokia |
| L. Granville | |
| Federal University of Rio Grande do Sul (UFRGS) | |
| J. Tantsura | |
| Apstra, Inc. | |
| November 4, 2019 |
Intent-Based Networking - Concepts and Overview
draft-clemm-nmrg-dist-intent-03
Intent and Intent-Based Networking are taking the industry by storm. At the same time, those terms are used loosely and often inconsistently, in many cases overlapping and confused with other concepts such as "policy". This document is intended to clarify the concept of "Intent" and provide an overview of functionality that associated with it. The goal is to contribute towards a common and shared understanding of terms, concepts, and functionality which can be used as foundation to guide further definition of associated research and engineering problems and their solutions.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on May 7, 2020.
Copyright (c) 2019 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
Traditionally in the IETF, interest with regard to management and operations has focused on individual network and device features. Standardization emphasis has generally been put on management instrumentation that needed to be provided to a networking device. A prime example for this is SNMP-based management and the 200+ MIBs that have been defined by the IETF over the years. More recent examples include YANG data model definitions for aspects such as interface configuration, ACL configuration, or Syslog configuration.
There is a sense and reality that in modern network environments managing networks by configuring myriads of "nerd knobs" on a device-by-device basis is no longer sustainable. Big challenges arise with keeping device configurations not only consistent across a network, but consistent with the needs of services and service features they are supposed to enable. Adoptability to changes at scale is a fundamental property of a well designed IBN system, that requires abilty to consume and process analytics that are context/intent aware at near real time speeds. At the same time, operations need to be streamlined and automated wherever possible to not only lower operational expenses, but allow for rapid reconfiguration of networks at sub-second time scales and to ensure networks are delivering their functionality as expected.
Accordingly, IETF has begun to address end-to-end management aspects that go beyond the realm of individual devices in isolation. Examples include the definition of YANG models for network topology [RFC8345] or the introduction of service models used by service orchestration systems and controllers [RFC8309]. In addition, a lot of interest has been fueled by the discussion about how to manage autonomic networks as discussed in the ANIMA working group. Autonomic networks are driven by the desire to lower operational expenses and make management of the network as a whole exceptionally easy, putting it at odds with the need to manage the network one device and one feature at a time. However, while autonomic networks are intended to exhibit "self-management" properties, they still require input from an operator or outside system to provide operational guidance and information about the goals, purposes, and service instances that the network is to serve.
This vision has since caught on with the industry in a big way, leading to a significant number solutions that offer "intent-based management" that promise network providers to manage networks holistically at a higher level of abstraction and as a system that happens to consist of interconnected components, as opposed to a set of independent devices (that happen to be interconnected). Those offerings include IBN systems (offering full lifecycle of intent), SDN controllers (offering a single point of control and administration for a network) as well as network management and Operations Support Systems (OSS).
However, it has been recognized for a long time that comprehensive management solutions cannot operate only at the level of individual devices and low-level configurations. In this sense, the vision of "intent" is not entirely new. In the past, ITU-T's model of a Telecommunications Management Network, TMN, introduced a set of management layers that defined a management hierarchy, consisting of network element, network, service, and business management. High-level operational objectives would propagate in top-down fashion from upper to lower layers. The associated abstraction hierarchy was key to decompose management complexity into separate areas of concerns. This abstraction hierarchy was accompanied by an information hierarchy that concerned itself at the lowest level with device-specific information, but that would, at higher layers, include, for example, end-to-end service instances. Similarly, the concept of "policy-based management" has for a long time touted the ability to allow users to manage networks by specifying high-level management policies, with policy systems automatically "rendering" those policies, i.e. breaking them down into low-level configurations and control logic.
What has been missing, however, is putting these concepts into a more current context and updating it to account for current technology trends. This document attempts to clarify the concepts behind intent. It differentiates it from related concepts. It also provides an overview of first-order principles of Intent-Based Networking as well as associated functionality. In addition, a number of research challenges are highlighted. The goal is to contribute to a common and shared understanding that can be used as a foundation to articulate research and engineering problems in the area of Intent-Based Networking.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.
The following section provides an overview of the concept of intent respectively intent-based management. It also provides an overview of the related concepts of service models, and of policies respectively policy-based management, and explains how they relate to intent and intent-based management.
In the context of Autonomic Networks, Intent is defined as "an abstract, high-level policy used to operate a network" [RFC7575]. According to this definition, an intent is a specific type of policy. However, to avoid using "intent" simply as a synonym for "policy, a clearer distinction needs to be introduced that distinguishes intent clearly from other types of policies.
For one, while Intent-Based Management clearly aims to lead towards networks that are dramatically simpler to manage and operate requiring only minimal outside intervention, the concept of "intent" is not limited to autonomic networks, but applies to any network. Networks, even when considered "autonomic", are not clairvoyant and have no way of automatically knowing particular operational goals nor what instances of networking services to support. In other words, they do not know what the "intent" of the network provider is that gives the network the purpose of its being. This still needs to be communicated by what informally constitutes "intent".
More specifically, intent is a declaration of operational goals that a network should meet and outcomes that the network is supposed to deliver, without specifying how to achieve them. Those goals and outcomes are defined in a manner that is purely declarative - they specify what to accomplish, not how to achieve it. "Intent" thus applies several important concepts simultaneously:
In an autonomic network, intent should be rendered by the network itself, i.e. translated into device-specific rules and courses of action. Ideally, it should not even be orchestrated or broken down by a higher-level, centralized system, but by the network devices themselves using a combination of distributed algorithms and local device abstraction. Because intent holds for the network as a whole, not individual devices, it needs to be automatically disseminated across all devices in the network, which can themselves decide whether they need to act on it. This facilitates management even further, since it obviates the need for a higher-layer system to break down and decompose higher-level intent, and because there is no need to even discover and maintain an inventory of the network to be able to manage it.
Tentative definition for intent-based networks Networks configuring and adapting autonomously to the user or operator intentions (i.e., a desired state or behavior) without the need to specify every technical detail of the process and operations to achieve it (i.e., the "machines" will figure out on their own how to realize the user goal).
Other definitions of intent exist such as [TR523] and will be investigated in future revisions of this document. Likewise, some definitions of intent allow for the presence of a centralized function that renders the intent into lower-level policies or instructions and orchestrates them across the network. While to the end user the concept of "intent" appears the same regardless of its method of rendering, this interpretation opens a slippery slope of how to clearly distinguish "intent" from other higher-layer abstractions. Again, these notions will be further investigated in future revisions of this document and in collaboration with NMRG.
A service model is a model that represents a service that is provided by a network to a user. Per [RFC8309], a service model describes a service and its parameters in a portable/vendor agnostic way that can be used independent of the equipment and operating environment on which the service is realized. Two subcategories are distinguished: a "Customer Service Model" describes an instance of a service as provided to a customer, possibly associated with a service order. A "Service Delivery Model" describes how a service is instantiated over existing networking infrastructure.
An example of a service could be a Layer 3 VPN service [RFC8299], a Network Slice, or residential Internet access. Service models represent service instances as entities in their own right. Services have their own parameters, actions, and lifecycles. Typically, service instances can be bound to end users, who might be billed for the service.
Instantiating a service typically involves multiple aspects:
They involve a system, such as a controller, that provides provisioning logic. Orchestration itself is generally conducted using a "push" model, in which the controller/manager initiates the operations as required, pushing down the specific configurations to the device. (In addition to instantiating and creating new instances of a service, updating, modifying, and decommissioning services need to be also supported.) The device itself typically remains agnostic to the service or the fact that its resources or configurations are part of a service/concept at a higher layer.
Instantiated service models map to instantiated lower-layer network and device models. Examples include instances of paths, or instances of specific port configurations. The service model typically also models dependencies and layering of services over lower-layer networking resources that are used to provide services. This facilitates management by allowing to follow dependencies for troubleshooting activities, to perform impact analysis in which events in the network are assessed regarding their impact on services and customers. Services are typically orchestrated and provisioned top-to-bottom, which also facilitates keeping track of the assignment of network resources. Service models might also be associated with other data that does not concern the network but provides business context. This includes things such as customer data (such as billing information), service orders and service catalogues, tariffs, service contracts, and Service Level Agreements (SLAs) including contractual agreements regarding remediation actions.
Like intent, service models provide higher layers of abstraction. Service models are often also complemented with mappings that capture dependencies between service and device or network configurations. Unlike intent, service models do not allow to define a desired "outcome" that would be automatically maintained by the intent system. Instead, management of service models requires development of sophisticated algorithms and control logic by network providers or system integrators.
Policy-based management (PBM) is a management paradigm that separates the rules that govern the behavior of a system from the functionality of the system. It promises to reduce maintenance costs of information and communication systems while improving flexibility and runtime adaptability. It is present today at the heart of a multitude of management architectures and paradigms including SLA-driven, Business-driven, autonomous, adaptive, and self-* management [Boutaba07]. The interested reader is asked to refer to the rich set of existing literature which includes this and many other references. In the following, we will only provide a much-abridged and distilled overview.
At the heart of policy-based management is the concept of a policy. Multiple definitions of policy exist: "Policies are rules governing the choices in behavior of a system" [Sloman94]. "Policy is a set of rules that are used to manage and control the changing and/or maintaining of the state of one or more managed objects" [Strassner03]. Common to most definitions is the definition of a policy as a "rule". Typically, the definition of a rule consists of an event (whose occurrence triggers a rule), a set of conditions (that get assessed and that must be true before any actions are actually "fired"), and finally a set of one or more actions that are carried out when the condition holds.
Policy-based management can be considered an imperative management paradigm: Policies specify precisely what needs to be done when and in which circumstance. Using policies, management can in effect be defined as a set of simple control loops. This makes policy-based management a suitable technology to implement autonomic behavior that can exhibit self-* management properties including self-configuration, self-healing, self-optimization, and self-protection. In effect, policies define management as a set of simple control loops.
Policies typically involve a certain degree of abstraction in order to cope with heterogeneity of networking devices. Rather than having a device-specific policy that defines events, conditions, and actions in terms of device-specific commands, parameters, and data models, policy is defined at a higher-level of abstraction involving a canonical model of systems and devices to which the policy is to be applied. A policy agent on a controller or the device subsequently "renders" the policy, i.e., translates the canonical model into a device-specific representation. This concept allows to apply the same policy across a wide range of devices without needing to define multiple variants. In other words - policy definition is de-coupled from policy instantiation and policy enforcement. This enables operational scale and allows network operators and authors of policies to think in higher terms of abstraction than device specifics and be able to reuse the same, high level definition defintion across different networking domains, WAN, DC or public cloud.
Policy-based management is typically "push-based": Policies are pushed onto devices where they are rendered and enforced. The push operations are conducted by a manager or controller, which is responsible for deploying policies across the network and monitor their proper operation. That said, other policy architectures are possible. For example, policy-based management can also include a pull-component in which the decision regarding which action to take is delegated to a so-called Policy Decision Point (PDP). This PDP can reside outside the managed device itself and has typically global visibility and context with which to make policy decisions. Whenever a network device observes an event that is associated with a policy, but lacks the full definition of the policy or the ability to reach a conclusion regarding the expected action, it reaches out to the PDP for a decision (reached, for example, by deciding on an action based on various conditions). Subsequently, the device carries out the decision as returned by the PDP - the device "enforces" the policy and hence acts as a PEP (Policy Enforcement Point). Either way, PBM architectures typically involve a central component from which policies are deployed across the network, and/or policy decisions served.
Like Intent, policies provide a higher layer of abstraction. Policy systems are also able to capture dynamic aspects of the system under management through specification of rules that allow to define various triggers for certain courses of actions. Unlike intent, the definition of those rules (and courses of actions) still needs to be articulated by users. Since the intent is unknown, conflict resolution within or between policies requires interactions with a user or some kind of logic that resides outside of PBM.
What Intent, Policy, and Service Models all have in common is the fact that they involve a higher-layer of abstraction of a network that does not involve device-specifics, that generally transcends individual devices, and that makes the network easier to manage for applications and human users compared to having to manage the network one device at a time. Beyond that, differences emerge. Service models have less in common with policy and intent than policy and intent do with each other.
Summarized differences:
One analogy to capture the difference between policy and intent systems is that of Expert Systems and Learning Systems in the field of Artificial Intelligence. Expert Systems operate on knowledge bases with rules that are supplied by users. They are able to make automatic inferences based on those rules, but are not able to "learn" on their own. Learning Systems (popularized by deep learning and neural networks), on the other hand, are able to learn without depending on user programming. However, they do require a learning or training phase and explanations of actions that the system actually takes provide a different set of challenges.
The following operating principles allow characterizing the intent-based/-driven/-defined nature of a system.
Additional principles will be described in future revision of this document addressing aspects such as: Target groups not individual devices, agnostic to implementation details, user-friendly, user vocabulary vs. language of the device/network, explainability, validation and troubleshooting, how to resolve and point out conflicts (between intents), reconcile the reality of what is possible with the fiction of what the user would want, "moderate", awareness of operating within system boundaries, outcome-driven ((what not how, for the user);(what and how/where, for the operator).not imperative/instruction based.).
The above principles will be further used to understand implications on the design of intent-based systems and their supporting architecture, and derive functional and operational requirements.
Intent is subject to a lifecycle: it comes into being, may undergo changes over the course of time, and may at some point be retracted. This lifecycle is closely tied to various interconnection functions that are associated with the intent concept.
Figure 1 depicts an intent lifecycle and its main functions. The functions are divided into two functional (horizontal) planes and into three (vertical) spaces.
The functional planes provide structure for the main funcional concerns that are associated with intent: how to fulfill intent, and how to assure it.
The spaces indicate the different perspectives and interactions with different roles that are involved to address the functions:
User Space : Translation / IBS : Network Ops
: Space : Space
: :
+---------+ : +----------+ +-----------+ : +---------+
Fulfill |recognize| ---> |translate/|-->|learn/plan/| ---> | config/ |
|intent | <--- | refine | | render | : |provision|
+----^----+ : +----------+ +-----^-----+ : +---------+
| : | : |
.............|................................|..................|.....
| : +----+---+ : v
| : |validate| : +----------+
| : +----^---+ <------| monitor/ |
Assure +---+---+ : +---------+ +-----+---+ : | observe/ |
|report | <---- |abstract |<---| analyze | <------| assure |
+-------+ : +---------+ |aggregate| : +----------+
: +---------+ :
Figure 1: Intent Lifecycle
When inspecting the diagram carefully, it become apparent that the intent lifecycle in fact involves two cycles, or loops:
Slight alternatives in intent lifecycles and the functions involved are conceivable. Figure 2 depicts one such alternative with an emphasis in intent fulfilment. (Todo: Intent attributes, intent states. Distinguish flow from users to network, and from network to user.)
user related user data <-----<-+--------+
data + + | |
| | | |
+----v------+ +-----v-----+ | |
| recognize +---+ +-----+ generate | | |
user +-----------+ | | +-----------+ | |
space | | | |
+--------------------------------------------------------------------+
system | | | |
space +---v---v---+ +----------+ +-----+-----+ |
| translate <-->+ validate <---> recommend | |
+-----+-----+ +----------+ +-----------+ |
| |
+-----v-----+ |
| normalize | |
+-----+-----+ |
| |
+-----v-----+ |
| decompose | |
+-----+-----+ |
| |
+------v------+ |
| communicate | |
+------+------+ |
preparation | |
phase | |
+-------------------------------------------------------------------+
operation | |
phase +-----v----+ |
| fullfill | |
+-----+----+ |
| |
+----v----+ +--------+ |
| observe +-----> report +-------------------+
+----+----+ +--------+
|
+----v---+
| assure |
+--------+
Figure 2: Intent Lifecycle (alt.)
Intent-Based Networking involves a wide variety of functions which can be roughly divided into two categories:
The following sections provide a more comprehensive overview of those functions.
RBD
Ability to reason about system' state by employing closed-loop validation in the presence of an inevitable change is a fundamental property of an Intent Assurance part of an IBN system. Since service expectations are created during intent consumption and modeling phase, closed-loop intent vaidation should start immidiatelly, with the service instantiation. Telemetry consumed could then be enriched with an additional context and must always be processed in context of the Intent it has been instantiated. Direct relationship between the Intent and telemetry gathered enables correlation between changes in states and the Intent and provides contextual base for reasoning about the changes.
Arguably, given the popularity of the term intent, its use could be broadened to encompass also known concepts ("intent-washing"). For example, it is conceivable to introduce intent-based terms for various concepts that, although already known, are related to the context of intent. Each of those terms could then designate an intent subcategory, for example:
Whether to do so is an item for discussion by the Research Group.
Not applicable
Not applicable
| [RFC2119] | Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997. |
| [RFC8174] | Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017. |