On Thursday, 23 January 2020 18:59:19 UTC Karl Reuss wrote:
> The current definition of Instance in the RSSAC026 lexicon is somewhat
> ambiguous.
> > When anycast routing is used to allow more than one server to have the
> > same IP address, each one of those servers is commonly referred to as an
> > instance. For root servers, one refers to "an instance of J-root" to mean
> > one of the network locations answering to J-root’s IP address.
> The first sentence says it's a server, and the second sentence says it's a
> location. In the current draft, the definition has been changed to this:
> > *When anycast routing is used to allow more than one server to have the
> > same IP address, each one of those servers is commonly referred to as an
> > instance or anycast instance. For root servers, one refers to "an
> > instance of J-root" to mean one of the network servers answering to one
> > of J-root’s IP addresses.*
> So now it's a server on one address (which implies there's another instance
> on J-root's other IP address).
>
> Paul thought this could use some discussion and I agree.
i agree also.
> In my mind, using the word 'server' to define something in this context is
> confusing. Is a server a piece of equipment or a software process? I
> think of an instance as "equipment used at a location used by an operator
> to provide DNS service". If I had to define it as a single word, I would
> say it's a location.
since we've known here of cases where two instances in adjacent racks each had
multiple servers and its own local load balancer but spoke to two different
internet exchanges which were present in the same data center, "location" is
wrong. that's in fact where we got "instance".
> At one location for D-root we have 4 Dell servers in the same rack. They are
> each talking BGP with a router that's using ECMP to reach them. Each Dell
> is answering on v4 and v6. I consider this one instance.
>
> What do others think?
i agree, that's an "instance", because it is a dedicated collection of server
and network equipment operated as a root name server, having upstream
connectivity not shared by any other instance. however, a lot of rootops no
longer operate their own equipment in some "location(s)", and so share server
and network equipment with "instances" of other root servers. so, "instance"
will have to be defined to account for virtualization including both the
cloudflare model and the PCH model. i don't know yet exactly how to do that.
noting that c-root has always used dedicated server equipment but has never
had dedicated network equipment (it plugs into cogentco edge routers which may
also serve other customers), "instance" may have to be defined in terms of
unique servers which might be virtual, having unique upstream connectivity
among other instances of the same root server. by leaving out "equipment" and
leaving out "network", a version of "instance" might suit all current and
contemplated root server deployments, including the cloudflare and PCH models.
I think it may be a bit cleaner to discuss anycast (generally) and then what is meant by an anycast instance, and finally what an anycast instance of a particular root server identity is.
Here's a brief attempt at this:
Anycast is the routing technique of advertising the reachability of a prefix from more than one router, typically (but not exclusively) as a global BGP announcement.
An anycast routing instance is a single such routing announcement.
The anycast instance of a root server identity is the server or set of servers reachable via a single anycast routing instance of the corresponding identity's IPv4 and/or IPv6 prefix.
It leaves unspecified any of the (IMHO unimportant details) of what the servers actually are (physical/virtual, dedicated/shared), and leaves the "network location" as an abstracted "anycast routing instance", for similar reasons.
It also allows for flexibility in whether the servers are single-stack or dual-stack for IPv4/IPv6 announcements/services.
Brian