At sufficiently large operational scales, IP addressing becomes a drag on innovation in network- and web-oriented services. Before any architectural change — and certainly at the start of designing something new — the unavoidable questions arrive: which block of addresses can we use? Do we have enough IPv4, and if not, where do we get more? How do we use IPv6, and does that choice affect other IPv6 uses? And what planning, checks, time and people will migration demand? Each of these costs time, money and resources.
That IP addresses should be the last thing a network has to think about is precisely what their design intended, more than 40 years after the original protocol. But the Internet has shown that small weaknesses, often invisible at design time, surface at scale.
"More addresses" is not the answer. In IPv4 it only deepens scarcity and pushes market prices higher. IPv6 is necessary but only part of a solution: best practice there allocates a /56 even for personal use — 272, or roughly 4,722,000,000,000,000,000,000 addresses, a quantity nobody can reason about.
Addressing Agility is the approach described here, with full details in an ACM SIGCOMM 2021 paper. The central observation is simple: neither the number of names on a single address, nor an address changing with every query, nor changes made purely for provisioning, policy or performance reasons, has any hard limit.

The underlying insight: in IP, as in the global postal system, addresses have never been, and are never, needed to represent names — we merely treat them that way sometimes. All names should be able to share all of their addresses, any subset of them, or even just one.
The narrow waist as a choke point
Conventions decades old tie addresses to names and resources, a natural inheritance from a world where one computer had one name and typically one network interface card. For end clients and network carriers — little need for names, less for listening processes — the bindings barely matter. For content hosting, distribution and content-service providers (CSPs), they impose firm limits. Once attached to names, interfaces and sockets, addresses become largely static and any change requires effort, planning and care, if it is possible at all.
The IP narrow waist made the Internet possible, but as TCP did for transport and HTTP for application protocols, it has become a bottleneck to innovation. Separate communication bindings (with names) and connection bindings (with interfaces and sockets) produce transitive relationships between them.

The transitive lock resists breaking because changing either side affects the other; providers also use addresses to encode policies and service levels that exist independently of names. The bindings are one more thing to think about, for no good reason, and questions such as "which addresses do we use?" should never lead the design of a new architecture.
These bindings are not only artificial but, per the original standards, incorrect. RFC 791 is explicit that a distinction is made between names, addresses and routes: a name indicates what we seek, an address where it is, a route how to get there. Attaching information such as SNI or an HTTP host to an IP address in a higher-layer protocol violates the layering principle.
The path to decoupling
The industry has been pursuing addressing agility for roughly two decades. Google's Maglev broke the one-to-one binding between IP and network interface by combining Equal Cost MultiPath (ECMP) with consistent hashing to spread traffic from one 'virtual' IP across many servers — a use the original IP RFCs proscribe, and nothing about it is virtual. Similar systems followed at GitHub, Facebook and elsewhere, including Cloudflare's own Unimog. More recently Cloudflare built a programmable sockets architecture, bpf_sk_lookup, to decouple addresses from sockets and processes.
Names were the missing half. HTTP 1.1 made the host field mandatory in 1997, formally acknowledging that multiple names can share one address; TLS reproduced this with Server Name Indication. Both are absolute requirements, because names outnumber addresses.
If a name is reachable at an address, any address can reach it — the name-to-address mapping is any-to-any. DNS-based load-balancing has offered many addresses per name since 1995. The step further is any addresses at a given time for all names, or one address for all names.
Inverting the mapping: policies first, names ignored
Addressing agility rests on authoritative DNS, but not on static name-to-IP records. From a client's perspective the binding exists only on-query: the response is the last moment in a request's life when a name can be bound to an address. Mappings are therefore made at the moment the response is returned, not stored in a zone file. Conventional DNS does the reverse — it looks up the addresses for a name and then applies policy, such as service level or geo-region, to narrow the selection.


Addressing agility inverts this. The starting point is a policy that may or may not include a name; in our deployment the policy ignored the name and used attributes such as location and account type. Those attributes identify a pool of addresses, either isolated to one policy or shared across several, and every address in the pool is equivalent — any of them may be returned, even at random, without inspecting the query name. Two implications follow per query: addresses are computed and assigned at runtime, and the IP-to-name mapping lives for the larger of the connection lifetime and the downstream cache TTL. The binding is ephemeral and can change without regard to earlier bindings, resolvers, clients or purpose — and scale is not a constraint, as an edge deployment demonstrated.
IPv6 changes nothing here
Everything said about IPv4 applies equally to IPv6; addresses are addresses, and they carry the same relatively static nature. Does switching to IPv6 satisfy the reasons to pursue agility? No. IPv6 mitigates exhaustion for the lifetimes of everyone alive today, but abundance of prefixes and addresses makes reasoning about their bindings to names and resources harder. Operators can also embed meaning in the bit length and large prefixes, which risks waste: many addresses in any prefix go unused. Cloudflare remains one of IPv6's biggest advocates, not least because abundance ensures longevity, yet IPv6 changes little about how addresses are tied to names and resources, whereas agility provides flexibility across their lifetimes.
Agility is not only for hyperscalers
Addressing Agility is usable, and desirable, for any service that operates authoritative DNS. Other content-oriented providers are obvious candidates, but so are smaller operators — universities, enterprises and governments that run their own authoritative services. Any operator able to accept connections on the returned addresses can benefit.
Production traffic has been served under Addressing Agility at the edge since June 2020. The deployment covers more than 20 million hostnames and services, all data centers in Canada, and a /20 (4096 addresses) in IPv4 alongside a /44 in IPv6; from January to June 2021 the working set shrank to a /24 (256 addresses). For every query, a random host-portion is generated within the prefix.
The most extreme form of the test came next. In June 2021, first in Montreal and soon after in Toronto, all 20+ million zones were mapped to a single address — an arrangement referred to internally as Ao1. Over one year, each policy-covered query received an address selected at random from as few as 4096 addresses, then 256, then one.
Observing the rollout
Queries landing outside the deployment data centers return the same addresses for a zone, as Cloudflare's global anycast system normally does. A query that reaches a deployment data center instead receives a random address, visible across successive dig commands against two different data centers:

Request traffic follows accordingly: servers are configured to accept connection requests for any of the 20+ million domains on all addresses in the pool. No surrounding systems needed modification — the change is a drop-in transparent modification to the authoritative DNS data pipeline, and BGP prefix advertisements, DDoS, load balancers and the distributed cache all ran untouched.
Load distribution as a side effect
Randomization behaves toward IP addresses the way a good hash function behaves toward a hash table, mapping an arbitrary-sized input evenly onto a fixed number of outputs. Measurements of load per IP, taken from 1% request samples at one data center over seven days, show the effect.

Before Addressing Agility

Randomization on /20

Randomization on /24
Before randomization, across only a small portion of Cloudflare's IP space, the gap between greatest and least requests per IP spanned three orders of magnitude, and bytes per IP almost six. Randomized onto a single /20 that previously occupied multiple /20s, those gaps fell to 2 and 3 orders of magnitude; at /24, with 20+ million zones spread over 256 addresses, differences in load shrink to small constant factors. This is relevant to any content service provider that provisions resources by IP address, since predicting a customer's load in advance is hard. The evidence points toward giving all the addresses to all the names.
Where assumptions about IP meaning break
Address randomization exposes weaknesses in systems whose designers ascribe meaning to an IP address beyond reachability; such meaning is the apparent prerequisite for any system that randomization might affect. Addressing Agility preserves, and even restores, the semantics of IP addresses and the core Internet architecture, but software built on assumptions about that meaning will fail. Each of these weaknesses is incidentally avoided by Ao1.
- HTTP connection coalescing. Clients such as Firefox that permit coalescing when the URI authority matches the connection are unaffected. Clients that require a URI host to resolve to the same IP as the connection will fail.
- Non-TLS or HTTP-based services.
ssh, for example, keeps a hostname-to-IP mapping inknown_hosts. That association is outdated and already broken by DNS records returning multiple addresses. - Non-SNI TLS certificates need a dedicated IP address, which forces providers to charge a premium since one address supports only one certificate. The larger issue, independent of IP, is TLS without SNI; efforts are underway to understand non-SNI and end the legacy.
- DDoS protections keyed on destination IP may be hindered initially. Randomization distributes attack load across all addresses in use, acting as a layer-3 load-balancer, and DoS mitigations commonly work by changing IP addresses — an ability Addressing Agility provides inherently.
Reducing to a single address
If randomization works over n addresses, why not over 1? A well-designed hash-based structure keeps its properties at any size, even 1, which makes this the true test of the foundations. The address set went from /20 to /24 and then, from June 2021, to one /32 (equivalently a /128). It doesn't just work — it really works. Non-TLS and non-HTTP services gain a reliable IP address, at least until a policy change on the name. HTTP connection coalescing falls out for free, and higher coalescing levels are observed where Ao1 is used.
For IPv6, the argument that a single address is unnecessary because exhaustion is unlikely is a pre-CIDR position — benign at best and irresponsible at worst. IPv6's address volume makes reasoning about addresses difficult; instead of asking why bind to one, the right question is "why not?"
Upstream effects
Ao1 surfaces a different set of implications, offering a window into the future of routing and reachability by amplifying the effects of small actions. Variable-length names in the universe will always outnumber fixed-length addresses, so by the pigeonhole principle single IP addresses must be shared by multiple names and by content from unrelated parties. So far none of the effects below have appeared in evaluations or in communications with upstream networks.
- Upstream routing errors become immediate and total. If all traffic arrives on one address or prefix, an upstream routing error hits all content equally. This is why Cloudflare returns two addresses in non-contiguous ranges — though the same applies to threat blocking.
- Upstream DoS protections could trigger. Concentrated traffic on a single address could be perceived upstream as an attack and trip existing protections.
Both are mitigated by Addressing Agility's ability to change addresses en masse quickly; prevention requires open communication and discourse. One further effect:
- IPv4 NAT port exhaustion may accelerate, and IPv6 solves it. On the client side, concurrent connections to one address are capped by the transport protocol's port field — roughly 65K for TCP. Linux addressed this recently (see the kernel commit and
SO_BIND_ADDRESS_NO_PORTinip(7)); in UDP the issue remains; for QUIC, connection identifiers can prevent exhaustion but must actually be used. No evidence of the problem has appeared so far.
To the best of current knowledge, this is the only risk attached to one-address use, and it disappears on migration to IPv6.
Open problems
With no limit on names per IP and per-query address changes available for any reason, the design space is just opening: BGP route leak detection and mitigation for anycast systems, measurement platforms and more are planned. Technical detail and acknowledgements appear in the accompanying paper and short talk. Among the open questions:
- What policies can be reasonably expressed or implemented?
- Is there an abstract syntax or grammar for expressing them?
- Could formal methods and verification prevent erroneous or conflicting policies?




