The IPFS stack is a suite of specifications and tools that share two key characteristics:
This document provides context and details about these characteristics. In doing so it defines what is or is not an IPFS implementation. This is a living document; it is expected to change over time as we define more of the principles that guide the architecture of IPFS or find clearer ways of describing those we have already defined.
The web's early designers conceived it as a universal space in which identifiers map to information resources. As the web grew, they enshrined in web architecture that all resources should have an identifier and defined "addressability" as meaning that "a URI alone is sufficient for an agent to carry out a particular type of interaction." ([webarch])
This design is tremendously successful. For all its flaws, the web brings together a huge diversity of software, services, and resources under universal addressability.
Unfortunately, HTTP addressability is based on a hierarchy of authorities that places resources under the control of a host and places hosts under the control of the DNS system (further issues with this model are discussed further in the Appendix). As indicated in [RFC3986]:
Many URI schemes include a hierarchical element for a naming authority so that governance of the name space defined by the remainder of the URI is delegated to that authority (which may, in turn, delegate it further).
CIDs in IPFS offer an improvement over HTTP URLs by maintaining universal addressability while eliminating the attack vectors inherent in hierarchical authority. Content addressability derives identifiers from the content of an information resource, such that any party can both mint the identifier and verify that it maps to the right resource. This eliminates the need for any authority outside of the resource itself to certify its content. It makes CIDs the universal self-certifying addressability component of the web.
Addressing data using CIDs is the first defining characteristic of IPFS. And the second characteristic, transport-agnosticity, can be supported thanks to the verifiability that CIDs offer. Across a vast diversity of implementations, architectures, and services, IPFS is the space of resources that can be interacted with over arbitrary transports using a CID. As Juan Benet once put it, "That's it!"
Conversely, any system that exposes interactions with resources based on CIDs is an IPFS implementation. There are many contexts in which CIDs can be used for addressing and content routing delegation ([http-routing-v1]) can support a wealth of interaction options by resolving CIDs.
Common wisdom about network protocol design is captured by Postel's Law or the Robustness Principle. Over the years it has developed multiple formulations, but the canonical one from [RFC1958] ("Architectural Principles of the Internet") is:
Be strict when sending and tolerant when receiving.
This principle is elegant, and expresses an intuitively pleasing behavior of protocol implementations. However, over the years, the experience of internet and web protocol designers has been that this principle can have detrimental effects on interoperability. As discussed in the Internet Architecture Board's recent work on Maintaining Robust Protocols, implementations that silently accept faulty input can lead to interoperability defects accumulating over time, leading the overall protocol ecosystem to decay.
There are two equilibrium points for protocol ecosystems: when deployed implementations are strict, new implementations, out of necessity, are required to be strict as well, leading to a strict ecosystem; conversely, when deployed implementations are tolerant, new implementations will have a strong incentive to tolerate non-compliance so as to interoperate. Tolerance is highly desirable for extensibility and adaptability to new environments, but strictness is highly desirable to prevent a protocol ecosystem from decaying into a complex collection of corner cases with poor or difficult interoperability (what the IETF refers to as "virtuous intolerance").
IPFS approaches this problem space with a new iteration on the robustness principle:
Be strict about the outcomes, be tolerant about the methods.
CIDs enforce strict outcomes because the mapping from address to content is verified; there is no room for outcomes that deviate from the intent expressed in an address. This strictness is complemented by a design that proactively expects change thanks to a self-describing format (CIDs are a multiformat implementation and support an open-ended list of hashes, codecs, etc.). The endpoints being enforceably strict means that everything else, notably transport, can be tolerant. Being tolerant about methods enables adaptability in how the protocol works, notably in how it can adapt to specific environments, and in how intelligence can be applied at the endpoints in novel ways, while being strict with outcomes guarantees that the result will be correct and interoperable.
Note that this approach to robustness also covers the End-to-end Principle. The end-to-end principle states that the reliability properties of a protocol have to be supported at its endpoints and not in intermediary nodes. For instance, you can best guarantee the confidentiality or authenticity of a message by encrypting or signing at one endpoint and decrypting or verifying at the other rather than asking relaying nodes to implement local protections. IPFS's aproach to robustness, via CIDs, is well aligned with that principle.
An IPFS Implementation:
These IPFS principles are broad. This is by design because, like HTTP, IPFS supports an open-ended set of use cases and is adaptable to a broad array of operating conditions. Considering cases at the boundary may help develop an intuition for the limits that these principles draw.
CIDs are readily made compatible with other content-addressable systems, but this does not
entail that all content-addressable systems are part of IPFS. Git's SHA1 hashes aren't CIDs
but can be converted into CIDs by prefixing them with
f01781114. Likewise, BitTorrent v2
uses multihashes in the
btmh: scheme. BitTorrent addresses aren't CIDs, but can be
converted to CIDs by replacing
The simplicity with which one can expose these existing system over IPFS by simply prefixing existing addresses to mint CIDs enables radical interoperability with other content-addressable systems.
The requirements above state that an implementation may forgo verification when "it is possible to ascertain that verification has happened elsewhere in a trusted part of the system." This is intended as a strict requirement in which implementors take trustlessness seriously, an indication that it's okay to not constantly spend cycles verifying hashes in an internal setup which you have reasons to believe is trustworthy. This is not a licence to trust an arbitrary data source just because you like them.
Authority is control over a given domain of competence. Naming authority is control over what resources are called.
Addressability is the property of a naming system such that its names are sufficient for an agent to interact with the resources being named.
Verifiability is the property of a naming system such that an agent can certify that the mapping between a name it uses and a resource it is interacting with is correct without recourse to an authority other than itself and the resource.
Self-certifying addressability is the property of a naming system such that it is both addressable and verifiable: any name is sufficient to interact with a resource, and its mapping to that resource can be certified without recourse to additional authority. Self-certifying addressability is a key component of a self-certifying web and it supports capture-resistance which can help mitigate against centralization.
CIDs support self-certifying addressability. With CIDs, the authority to name a resource resides only with that resource and derives directly from that resource's intrinsic property: its content. This frees interactions with CID-named resources from the power relation implicit in a client-server architecture. CIDs are the trust model of IPFS.
An implementation may retrieve a CID without verifying that the resource matches it, but that loses the resource's naming authority. Such an implementation would be comparable to an HTTP client looking DNS records up from a random person's resolver: it cannot guarantee that the addressing is authoritative. Implementers may make informed decisions as to where in their systems they support verification, but they should ensure that the mapping between CID and resource is verified whenever they have access to both the resource and the CID that maps to it.
We tend not to think about addressability because it is so foundational that we struggle to apprehend a system without it, but that is precisely why it is important that we get it right. You can find extensive historical evidence that TimBL and others saw URLs as arguably the most fundamental invention of the Web, and the early groups that worked on Web architecture discussed and debated the properties of URLs at length. The problems of centralization we face today trace their lineage back to those decisions.
The hierarchical nature of the HTTP addresses was intentional, as TimBL wrote clearly in Web Architecture from 50,000 feet:
The HTTP space consists of two parts, one hierarchically delegated, for which the Domain Name System is used, and the second an opaque string whose significance is locally defined by the authority owning the domain name.
The model that the Web's earlier designers had in mind was a federated model in which authority is delegated and addresses are owned based on that authority delegation. This is notably clear in the URI Ownership passage of the Architecture of the World Wide Web, Volume One:
URI ownership is a relation between a URI and a social entity, such as a person, organization, or specification. URI ownership gives the relevant social entity certain rights, including:
- to pass on ownership of some or all owned URIs to another owner—delegation; and
- to associate a resource with an owned URI—URI allocation.
By social convention, URI ownership is delegated from the IANA URI scheme registry, itself a social entity, to IANA-registered URI scheme specifications.(…)
The approach taken for the "http" URI scheme, for example, follows the pattern whereby the Internet community delegates authority, via the IANA URI scheme registry and the DNS, over a set of URIs with a common prefix to one particular owner. One consequence of this approach is the Web's heavy reliance on the central DNS registry.(…)
URI owners are responsible for avoiding the assignment of equivalent URIs to multiple resources. Thus, if a URI scheme specification does provide for the delegation of individual or organized sets of URIs, it should take pains to ensure that ownership ultimately resides in the hands of a single social entity. Allowing multiple owners increases the likelihood of URI collisions.
URI owners may organize or deploy infrastruture [sic] to ensure that representations of associated resources are available and, where appropriate, interaction with the resource is possible through the exchange of representations. There are social expectations for responsible representation management (§3.5) by URI owners. Additional social implications of URI ownership are not discussed here.
This notion of address or name ownership is pervasive across architectural documents. This passage from an interview of TimBL (Philosophical Engineering and Ownerhip of URIs) is explicit:
Alexandre Monnin: Regarding names and URIs, a URI is not precisely a philosophical concept, it's an artifiact [sic]. So you can own a URI while you cannot own a philosophical name. The difference is entirely in this respect.
Tim Berners-Lee: For your definition of a philosophical name, you cannot own it. Maybe in your world, in your philosophy, you don't deal with names that are owned, but in the world we're talking about, names are owned.
This expectation of delegated naming authority was so strong among early Web architects
that the development of naming conventions in HTTP space (eg.
.well-known paths) is described as "expropriation" in the
Web Architecture and the W3C's Technical Architecture
Group (TAG) issue on the topic stated that it "breaks the web".
Federated models only have weak capture-resistance because the federated entities can always concede power (precisely because they have ownership) but lack established means to support collective organization. As a result, any power imbalance will likely become hard to dislodge. A good example is search: as a publisher (the owner of delegated authority over your domain) you can cede the rights to index your content but you can't have a voice in what is done with the indexed content (individual opt out is not an option). This was fine when you could barter content for links, but once search power consolidated, the terms of trade deteriorated with no immediate recourse.
Many thanks to the following people, listed alphabetically, whose feedback was instrumental in producing this document: Adin Schmahmann, biglep, Dietrich Ayala, Juan Benet, lidel, Molly Mackinlay, and mosh.