Re: Alt-SvcB

[David Schinazi <[e-mail]>]

We'll be ready to drop support for Alt-Svc once browsers have access to
HTTPS RR everywhere. Until that happens, browsers will still use Alt-Svc
for new HTTP versions, and we will still add goodies to Alt-Svc. If the
IETF doesn't standardize those, then browsers will come up with their own
solutions. The problem here is once again misaligned incentives: browsers
want to be fast and to use new standards but it's the OSes that need to
change their DNS APIs. Do we know anyone at POSIX interested in revamping
getaddrinfo?

David

On Tue, Oct 25, 2022 at 2:25 PM Lucas Pardue <[e-mail]>
wrote:

> I'd also add that by deprecating Alt-Svc, we can reasonably descope it
> from other related work, to the advantage if the community.
>
> For instance, Martin Duke and I have a draft intended to optimise
> performance of selecting QUIC versions with HTTP/3 [1]. We describe how
> this works with the HTTPS record. We also have to accommodate Alt-Svc
> since  "that's what people use", even though we know that Alt-Svc's design
> is broken for this type of versioning support in certain scenarios like
> multi-CDN.
>
> I can live with contuinuing to send Alt-Svc in order to support legacy
> clients. Alt-Svc being frozen in time and not getting new goodies seems
> fair.
>
> Cheers
> Lucas
>
> [1] -
> https://datatracker.ietf.org/doc/draft-duke-httpbis-quic-version-alt-svc/
>
> On Tue, 25 Oct 2022, 21:33 Tommy Pauly, <[e-mail]> wrote:
>
>> The way I’d look at this is that we should be fine keeping the use of
>> Alt-Svc for existing (and what will become legacy) clients to upgrade to
>> h3, but we should not use it for any new protocol discovery. I.e., when we
>> have an HTTP version that needs some transport other than TCP and QUIC, we
>> shouldn’t plan on using Alt-Svc for that. So, our timeline should be to
>> make sure clients can do HTTPS RRs by the time we replace QUIC, which
>> should give us time.
>>
>> Tommy
>>
>> On Oct 25, 2022, at 1:21 PM, Ian Swett <[e-mail]> wrote:
>>
>> I would second David's statement.  In the world we live in today, we
>> still need to use the Alt-Svc header for a substantial number of users.
>>
>> On Tue, Oct 25, 2022 at 2:31 PM David Schinazi <[e-mail]>
>> wrote:
>>
>>> Hi Martin,
>>>
>>> Thanks for writing this up. Overall I think the long-term strategy makes
>>> sense, but I think it's too early to obsolete/deprecate 7838. It's fairly
>>> common for browsers to use getaddrinfo() on some platforms and that does
>>> not provide access to HTTPS RRs. In those cases, 7838 is the only path to
>>> using HTTP/3, so I expect browsers to keep using it for quite some time.
>>> Marking 7838 as obsolete doesn't reflect that reality.
>>>
>>> David
>>>
>>> On Mon, Oct 24, 2022 at 5:10 PM Martin Thomson <[e-mail]>
>>> wrote:
>>>
>>>> Hey everyone,
>>>>
>>>> The Alt-Svc design team has been very busy recently and making some
>>>> progress on working out an alternative alternative services design.
>>>>
>>>> I just posted
>>>> https://martinthomson.github.io/alt-svcb/draft-thomson-httpbis-alt-svcb.html
>>>> as a -00 draft.  This outlines the alternative design that we've been
>>>> exploring in the design team.
>>>>
>>>> The basic idea is split into two procedures:
>>>>
>>>> 1. Use: When an Alt-SvcB field or ALTSVCB frame is encountered, the
>>>> client looks for HTTPS records for the provided name in the DNS and creates
>>>> a connection using what it learns.
>>>> 2. Reuse: When a client that has previously used an alternative service
>>>> connects again, it remembers the HTTPS record that worked.  It performs a
>>>> regular HTTPS record lookup for the server - not using the alternative that
>>>> it learned, but the name from the URI - but it prefers the alternative it
>>>> previously used if that alternative appears in the results.
>>>>
>>>> The draft explains in more detail and goes into some of the
>>>> implications of the design.
>>>>
>>>> This is not done by any imagining.  We have a bunch of open issues at
>>>> https://github.com/martinthomson/alt-svcb/issues that do require some
>>>> amount of input.  But we think that this is a promising approach and would
>>>> appreciate more input.
>>>>
>>>> Cheers,
>>>> Martin
>>>>
>>>>
>>