[RFC] Using project and service specific DNS end points to support flexible deployments.

DJ Delorie dj@redhat.com
Tue Nov 10 01:39:27 GMT 2026


On Sat, Nov 7, 2026 at 10:51 AM Carlos O'Donell <carlos@redhat.com> wrote:
>
> On 11/6/26 1:06 PM, DJ Delorie wrote:
> > On Fri, Nov 6, 2026 at 11:10 AM Carlos O'Donell <carlos@redhat.com> wrote:
> >> Would you agree that doing {service, project} x DNS end points is orthogonal
> >> to the ingress scaling point you made?
> >>
> >> That is to say that we can always do ingress scaling.
> >>
> >> We should also do the DNS changes to allow moving services around easily.
> >
> > All I mean is that if we do "project.host/service" we have one
> > ingress/dns path, and we can use internal reverse proxies to manage
> > service load and per-service dns without the users even knowing (aside
> > from removing an ingress point).
>
> This...
>
> > With service.project.host/ you need an ingress concentrator for each
> > service, and service dns changes are public-facing and thus affect
> > users and need to be scheduled much further in advance.  Plus you need
> > to manage far more public DNS entries - number of services times
> > number of ingress points.
> ... and this are the same.
>
> The use of service.project.host does not *require* a separate ingress
> concentrator per service, but it permits it.
>
> In path-style the route is keyed at the reverse proxy by URL path.
>
> In subdomain-style the route is keyed at the reverse proxy by HTTP Host header.
>
> You can handle both in the same reverse proxy AFAIK.
>
> (1) Multiple ingress concentrators?
>
> No required, but you can split them to increase robustness.
>
> (2) Manage more public DNS entries?
>
> A wildcard record *.glibc.sourceware.org is one DNS entry.
> A wildcard cert *.glibc.sourceware.org (DNS-01 ACME) covers certs for all servies.
>
> > The only real advantage of the per-service DNS style is services that
> > can't be https proxied, like smtp or git.
>
> Origin isolation for anything browser based puts cookies in different
> security domains. If you don't do this then a bad services that has
> an XSS can read another's cookies and use those credentials to do
> something with a service that matters *more* than one of the side
> services.
>
> > But I suppose a hybrid approach can be used -
> > https://project.host/service for anything that's an https url, and
> > service.project.host for anything that can't be reverse proxied.
>
> You are not wrong, a hybrid approach might be good, I hadn't entirely
> reviewed how or where we would split the hybrid services like you suggest.
>
> In summary:
> - You do not *require* separate ingress concentrators.
> - Multiple DNS entry costs are mitigated (discussed in the RFC).
> - Some protos cant' be HTTP-proxied.
> - You'll want this later for IP/provider/GeoDNS/AnyCast which path style can't do (without a new hostname anyway)
> - You do expose users to public hostname changes (more coordination)
> - You gain the benefit of origin isolation in web-based services.
>
> --
> Cheers,
> Carlos.
>



More information about the Libc-alpha mailing list