[DNSOP] a CDN perspective on ANAME challenges

[Erik Nygren <[e-mail]>]

One of the intended goals of ANAME is to improve interoperability of
onboarding onto CDNs for URLs at a zone apex, such as “http(s)://example.com
”.

The TL;DR is that ANAME is unlikely to allow interoperability here unless
authorities are willing to effectively and scalablely do recursion-with-ECS
for all requests (caching keyed by ECS prefix and obeying short, sub-minute
TTLs).  Simply resolving “sibling address records” on a zone master (per
draft-ietf-dnsop-aname-04) is likely to be declared incompatible and
unsupported (or at least severely restricted) by at least some (if not
many) CDNs.

For some background, Dan York highlights some sample use-cases here:


https://tools.ietf.org/html/draft-york-dnsop-cname-at-apex-publisher-view-01

For example, web publishers want to do the equivalent of:

     example.com  300 IN CNAME a123qk.example-cdn.net.


Two use-cases (quoted from draft-york above):

1.  users will be able to simply enter "example.com" into their browser; and

2.  users will only see "example.com" in their address bar (if URLs are
even displayed).

Note that for some CDNs, the “a123qk.example-cdn.net” target name
dynamically selects which IP addresses to return based on a wide variety of
factors (recursive nameserver IP, end-user IP prefix passed via ECS,
dynamic load and liveness conditions, product, customer security
requirements, TLS certificate needed for non-SNI-only configurations, etc)
and often have short TTLs (often in the range of 20s to 60s, although
sometimes up to 300s).

Some CDNs have had vertically-integrated solutions for onboarding zone apex
names for years, allowing example.com to hand back CDN IPs as long as the
zone is delegated to the CDN authorities.  The need for these to be
vertically integrated comes from the level of complexity involved in
calculating these answers.

Even these vertically integrated solutions may have limitations.  For
example, one early vertically-integrated solution had very similar
functionality to some of the ANAME proposals.  It only covered case #1
above, handing out a subset of CDN IPs (with limited performance). Along
with this was a correlated site HTTP configuration that would issue a
30[127] redirect to “www.example.com” which would in-turn use a normal set
of CDN IPs.  This doesn’t help for the “users will only see ‘example.com’
in their address bar” case due to the need to redirect to a hostname that
can fully CNAME.

More advanced vertically-integrated solutions that can also handle case #2
may involve high-volume feeds of proprietary dynamic state and
configuration being fed to the authoritative nameservers, allowing them to
construct answers on-the-fly.

It is unclear if third-party authoritative servers will be able to do
enough here to be compatible with CDNs, at least until we get to a point
where most/all recursive servers support ANAME sibling address substitution
(and there are huge numbers of these around the world run by a huge variety
of operators!).

For third-party authoritative servers wishing to collapse A/AAAA names into
a zone as signalled by an ANAME, some of the options include:


   1.

   Primary zone master does sibling address substitution:  Some CDNs will
   declare this to be strictly incompatible.  (ie, all traffic would be
   directed to a small set of IPs associated with where the Primary Master is
   located).  This would defeat the purpose of using a non-purely-anycast
   CDN.
   2.

   Zone secondaries do frequent sibling address substitution:  Some CDNs
   may still declare this to be incompatible. Traffic is still directed to a
   small subset of IPs associated with where the zone secondaries are
   located.  If the zone secondaries are widely deployed, some CDNs might be
   able to support this using anycast configurations, although that might
   limit to a subset of products or limit to a subset of CDN server locations
   and might have CDN SLA limitations.
   3.

   Authoritative resolvers do a mixture of the following, which is likely
   what would be needed to be compatible with CDNs that do DNS-based traffic
   direction:


   -

      Online / on-demand lookups using ECS, with response caching.
      (Possibly requiring collaboration with the CDN to be on an ECS ACL or
      allowlist.)
      -

      Online DNSSEC signing.
      -

      Honoring short (eg, 20s-60s) response TTLs.
      -

      Having the scale/capacity and attack resilience to be able to support
      this possibly-low-cache-hit-rate request workload.


It is only the third of these that might actually work well enough, but may
add a huge amount of complexity to authorities, requiring the authority to
also have many recursive functionality.  It is possible that some
authoritative operators are willing and able to do this to enable
interoperability. However, if this set of functionality is what is required
to make ANAME work well enough to achieve its goals, then we should be
clear on that. It would be a mistake to go through all of the effort of
rolling out ANAME only to find that CDNs declare it to be incompatible.

I’ll add the caveat that the HTTPSSVC proposal doesn’t fully or magically
solve this problem either.  It is likely that A/AAAA records will need to
live at the zone apex for a long time to come. If those are static
addresses, they’ll likely still need to do 30[127]-style redirects to “
www.example.com” names.  However HTTPSSVC leaves an option for “example.com”
requests (eg, those containing an Alt-Used request header) to not need to
return a redirect, allowing a somewhat more graceful transition as client
adoption picks up.

    Erik