Re: Request for HTTP WG adoption: draft-dunglas-mercure (Mercure Protocol)

To: Mark Nottingham <mnot@mnot.net>
Cc: ietf-http-wg@w3.org
Subject: Re: Mercure Protocol: spec stabilized for 1.0, refreshed
Internet-Draft

Hi Mark,

Thanks, that makes sense. I'll take you up on the offer!

So, I am withdrawing the adoption request I was about to send
to you and Tommy and I'll come back with a problem statement instead.

What I plan to write, before IETF 127 and in time for the list to pick it
apart first: use cases, requirements, and scope, argued without reference
to Mercure or any other specific solution. The subtlety between adjacent
use cases is the part I most want reviewed, so I would rather draw those
boundaries in the open than assert them. Concretely, the distinctions I
think matter are:

- whether the client subscribes to a resource or to a topic decoupled from
  any resource, which determines whether an intermediary can fan out at all
- whether the subscriber, the publisher, or an intermediary is the party
  that knows the authorization policy
- whether missed events must be recoverable after a disconnect, and at what
  cost to the server
- whether one connection has to carry many independent subscriptions, which
  is where the browser connection limits and the ad-hoc WebSocket protocols
  come from

I will include non-goals. Media delivery is one of them, which is most of
the answer on MoQ: it focuses on low-latency audio, video, and timed
metadata,
with relay and producer discovery explicitly out.

The space I am describing is text: JSON, HTML, XML payloads over
plain HTTP, reaching browsers through EventSource with no new transport, and
discovery is a required part of it rather than an excluded one. The relay
topology looks similar if you squint; the transport, payloads, deployment
model, and client story do not. I will make that case properly in the
document, and I would rather be argued out of it there than assume it.

I'll also survey what is deployed, including work that isn't mine:
WebSub, Web Push, the many ad-hoc WebSocket protocols, Braid-HTTP, and the
resource-centric proposals such as Rahul's PREP and Events Query. The
interesting question is which requirements each one does and does not meet,
and that is a better basis for scoping than my draft is. I wrote a first
pass at that comparison in 2023, before any of this round of discussion:

https://dunglas.dev/2023/11/mercure-braid-prep-news-about-subscribing-to-http-resource-updates/

Two things from it are worth stating up front, because they shape the scope
rather than the solution.

The first is a deployment constraint that I think is a requirement and not
an implementation detail: a large part of the deployed web cannot hold a
persistent connection per subscriber at the origin. PHP serves a large
share of sites, and the CGI, FastCGI, and serverless execution models all
make a long-lived connection at the application awkward, expensive, or
impossible; short proxy timeouts finish the job. Any design that requires
the origin to keep a connection open per subscriber excludes those stacks,
which is a good part of why the ad-hoc WebSocket services exist. Whether
that constraint belongs in scope is a fair question for the room, but it
should be asked explicitly rather than assumed away.

The second is that I don't think this space has a single winner, and I said
so before I had a reason to be diplomatic about it.

The deployed constraints and the right long-term answer are two different
questions. Something like Mercure is an answer for now: PHP, serverless and
EventSource aren't going away soon, and a hub is what makes push work on
those stacks today. That is not an argument that a hub is the right shape
for HTTP in the long run. The web also needs a first-class answer, where
subscribing to a resource is a native part of HTTP rather than bolted
alongside it, which is what Braid and PREP are aiming at. The conclusion of
that 2023 article was that these designs occupy different points and that
content negotiation between them is more plausible than convergence on one;
that is also how the now and the future coexist instead of blocking each
other. It is part of why a problem statement is more useful here than an
adoption call.

On change control, since you raised it: I would rather answer that honestly
in front of the group than pre-negotiate it. There is deployed code and
there are implementers who will not want churn, and I have no illusion that
handing the protocol to a WG leaves it unchanged. Whether that trade is
worth making depends a good deal on what the discussion concludes the
problem actually is, so I would like to have the discussion first.

Remote is likely for San Francisco, though I will confirm nearer the time.
Let me know what you need from me for the agenda and by when.

On IANA, understood on both counts. I will approach the designated experts
about the existing "mercure" registrations and leave the proposed new
registries until the stream question is settled.

Thanks for the steer, this is more useful than a yes or no would have been.

Kévin Dunglas
Les-Tilleuls.coop

On Thu, Aug 13, 2026 at 7:27 AM Mark Nottingham <mnot@mnot.net> wrote:

> Hi Kévin,
>
> As you point out, there's a lot of previous discussion in this area. When
> we consider taking on sizeable chunks of new work, it's good to understand
> the use cases, requirements, and scope (including what's *not* in scope) --
> especially in cases like this, where there's a fair amount of subtlety in
> the differences between adjacent use cases, and there are significant
> trade-offs here.
>
> Added to that, there's also similar work going on elsewhere; MoQ is
> delivering something that, if you squint just right, looks like pub/sub for
> the Web.
>
> Issuing a call for adoption -- whatever the venue -- isn't going to be
> helpful in untangling that. What I'd suggest instead is that you talk about
> your use cases, requirements, and proposed scope of work generally, as a
> problem space and (in the first instance) without reference to a particular
> solution. If we can get buy-in and interest in that, we can then dive into
> the proposal and figure out what the right place to work on it is
> (historically, HTTP extensions _can_ go in HTTPbis, but sometimes they're
> big enough to deserve their own working group, like MASQUE).
>
> It may even be that the independent stream is the best option for you --
> if there's wide deployment and folks don't want to change, giving change
> control to the IETF might be tricky. But some amount of community
> discussion might help make that decision.
>
> If that makes sense, we can give you some time at the San Francisco
> meeting (remote if you can't make it in person) to move that discussion
> forward.
>
> Cheers,
>
>
> P.S. Regarding IANA, the best place to start for the existing
> registrations is the relevant experts. It's probably best to figure out the
> stream you're on before worrying about establishment of registries.
>
>
> > On 13 Aug 2026, at 3:06 am, Kévin Dunglas <kevin@dunglas.fr> wrote:
> >
> > Hi Mark, Tommy,
> >
> > As you may have seen on the HTTP mailing list, we just finished a
> modernized version of Mercure (https://mercure.rocks), and I am writing
> with a concrete request regarding its potential standardization.
> >
> > Here are were you can find the new version of the spec :
> >
> > - Datatracker: https://datatracker.ietf.org/doc/draft-dunglas-mercure/
> > - Spec source (pinned):
> https://github.com/dunglas/mercure/blob/v1.0.0-alpha.3/spec/mercure.md
> >
> > The I-D was first submitted in October 2018 (draft-dunglas-mercure-00).
> > The protocol has been in production for roughly seven years across
> > multiple language ecosystems and web frameworks (Symfony, API Platform,
> > Laravel, Spiral, and others). It defines a hub-based publish/subscribe
> > model over HTTP and SSE, covering server-to-client real-time updates
> > that today are typically handled by ad-hoc WebSocket protocols.
> >
> > The draft was brought to the WG in 2020 and discussed again in the 2022
> > "Mnot's Pub/Sub for the Web" thread; neither round produced a clear
> > adoption decision. I have refreshed the Internet-Draft to reflect the
> > spec stabilized for 1.0 (reference implementation tagged
> > v1.0.0-alpha.3) and announced it on the list on 12 August 2026, with the
> > changes since -07 summarized there:
> > https://lists.w3.org/Archives/Public/ietf-http-wg/2026JulSep/0138.html
> >
> > I would like the HTTP WG to consider adoption of this work, or,
> > if adoption is not appropriate, to give me a clear answer on that so the
> > path forward is unambiguous.
> >
> > Also, the first response on the list suggested HTTPAPI as the better
> > home, on the grounds that Mercure does not modify or extend HTTP itself.
> > That is a fair reading and I have no attachment to the HTTP WG
> > specifically. My hesitation is that the HTTPAPI charter is scoped to HTTP
> > extensions relating to HTTP APIs, typically new header or trailer fields,
> > and excludes APIs for specific applications, so Mercure may sit outside
> > that charter too. Both groups are in WIT under the same AD, so I am
> > copying Mike here rather than opening a parallel thread.
> >
> > If you consider HTTPAPI (or any other group) the right home, that pointer
> > answers my request just as well as adoption here, and I will take it
> > there.
> >
> > If no group is interested, my intent is to publish through the
> Independent
> > Submissions Editor (independent stream). Tehre is one area where WG
> input would be
> > valuable either way: the draft's IANA section updates two existing
> > registrations (the "mercure" well-known URI and link relation type), adds
> > three OAuth Protected Resource Metadata members and one HTTP field name,
> > and establishes two Mercure-specific registries under Specification
> > Required. I would welcome the WG's view on those registry policies and on
> > the expert-review criteria the draft proposes.
> >
> > I am happy to present at an upcoming interim or IETF meeting if that
> > would help the WG reach a decision.
> >
> > Thanks for your time,
> > Kévin Dunglas
> > Les-Tilleuls.coop
> >
>
> --
> Mark Nottingham   https://mnot.net/
>
>

Received on Thursday, 13 August 2026 11:00:36 UTC