My comments and answers point by point:
1- TXT records, being "not efficient" are not a basis for a practical
solution.
2- With SRV records, load balancing and backup configuration can be static
(no need for an update protocol). This is simple and sufficient for many
small configurations. The point is important in my view: Yahoo, Google etc.
have so many users that the law of large numbers applies to them; they can
use crude and slowly varying parameters to balance traffic on the basis of
what they measure. But more ordinary servers cannot.
3- Even when dynamic load balancing configuration is desired, solutions that
don't depend on frequent DNS queries are preferable because more scalable
than those that do.
4- Yes IP addresses are needed with SRV records, as always. As indicated in
point 1 of the proposed spec, they can be provided to DNS clients in A RRs
contained in the same response as the SRV records.
5- In my understanding, the proposal is not a "big change" to wellknown
services. It is only a "upward compatible extension" to them (and a fairly
simple one): no DNS client is obliged to look at SRV records; no group of
servers offering a common service is obliged to have SRV records in the DNS;
but if a service server group has SRV records in the DNS (and the
corresponing name to IP address A records), then all DNS clients that are
upgraded perform weighted load balancing and/or backup without needing a
high frequency of DNS updates.
6- What is the added value? Efficient load balancing offered to small server
groups, even if they have relatively limited traffic, and no need for DNS
clients
that make frequent service accesses to also make frequent DNS queries.
Besides, the new feature is based on existing record types, the scope of
which
is only extended in a natural way.
This fits with the need for simplicity of concepts and orthogonality of
functions.
7- Exemple of application.
- A web service is offered by two servers s1 and s2, with s1 twice as
powerful as s2.
- A third server s3 is to be used only if s1 and s2 are down (backup).
- The server name used to call the service is xxx (an FQDN).
- IP addresses of servers are a1, a2 and a3 respectively.
- The DNS server for xxx is supposed to be upgraded (i.e. sends SRV records
that match the same queried name a A records, and also the A records that
match the SRV provided targets).
DNS RRs to be set are set as follows:
xxx A IN a1
xxx A IN a2
xxx A IN a3
xxx SRV IN 0 2 80 xxx1 (priority 0, weight 2)
xxx SRV IN 0 1 80 xxx2 (priority 0, weight 1)
xxx SRV IN 1 1 80 xxx3 (priority 1, weight not
significant)
xxx1 A IN a1
xxx2 A IN a2
xxx3 A IN a3
(1) A non upgraded DNS client queries for xxx (it has a web request to serve
and it has in its cache no information about xxx with an unexpired TTL).
(2) It receives all the above RRs.
(3) It interprets only the first ones, the only ones it recognizes, and acts
as
though there would be no SRV records. Note that the three xxx A records
can be ordered according to whatever feature applies to them (the
new feature doesn't conflict with existing such a features. They continue to
apply for DNS clients that don't interpret SRV records in responses to their
A queries.
(4) An upgraded DNS client queries for xxx .
(5) It receives all the above RRs.
(6) It sees that there ares SRV records in the response.
It also sees that no additional DNS query is necessary for getting the IP
addresses of targets xxx1, xxx2 and xxx3 (their IP adresses are provided).
(7) Its successive accesses to the xx service are then preformed such that
a1 is used twice as frequently as a2, the first one being chosen randomly.
If for one access neither a1 nor a2 answers, an access to a3is attempted.
All accesses are performed on the indicated port 80 (the usual one).
Clearer?
Rémi Després
----- Original Message -----
From: "Rémi Després" <
[email protected]>
Newsgroups: microsoft.public.win2000.dns
Sent: Saturday, March 06, 2004 3:10 PM
Subject: Re: SRV RRs support in Internet Explorer?
