Resolver behaviour with rrset

[Kevin Darcy]

Sebastian Castro Avila wrote:

>Hi all:
>
>AFAIK, when a resolver queries for a RR it will receive a "referal answer"  
>(I don't know anything, but better go here) or an "authoritative answer"  
>(the record you asked for has this answers).
>
>The server will return the RR set in some order (not important for the  
>question).
>
>Taking an answer from Barry Margolin about RR sets and Round Robin, that  
>says:
>
>=============================================
>The resolver remembers the response times of the servers, and prefers
>the one that has the best record.  However, the matching is fuzzy, so
>close times are considered equivalent, and it load balances among them
>(I think it's random selection rather than true round robin, but in the
>long run they should both produce equal load).
>=============================================
>
>What resolvers implements this behaviour?
>
Barry was using "resolver" in a technical sense to mean any mechanism 
that resolves DNS names. The BIND nameserver includes a "resolver" 
component, as do other DNS implementations, e.g. dnscache (from djbdns), 
Microsoft DNS, and so on. As far as I know, they all implement the 
round-trip-time logic, because it only makes sense, and because it was 
suggested by one of the original RFCs -- see Section 7.2 and 7.3 of RFC 
1035.

>How does it work if a server became unreachable?
>
That is interpreted as a very *bad* round-trip time, and that nameserver 
gets penalized so that it is used less frequently. It will still 
occasionally be queried, though, because it might recover.

>How often the resolver updates the response time for the whole list of  
>known servers?
>
On every query, I believe. It's a continuously-updated value.

>Where can I get more information about this "feature"?
>
I gave some RFC references above. Also, the algorithm is described 
vaguely in pages 30-31 of the _DNS_and_BIND_ O'Reilly book, Fourth 
Edition, as well as referred to in the discussion of the "topology" 
feature (which allows you to influence the algorithm slightly) on pages 
281-282.

                                                                         
                                             -Kevin