Ulrich Drepper wrote:
Simo Sorce wrote:
Ulrich, some people claim that glibc's getaddrinfo breaks IPv4 DNS round
robin as it reorders also IPv4 address, can you comment on this? In some
cases it is important.
It is not necessary (and quite wrong) to implement RR in the server.
It's an approximation at best anyway. In no way better than
randomization. And nothing stops programs from randomizing the results
returned by getaddrinfo before using them. Normal programs won't do
that but there is nothing which stops programs like ntpd from doing this.
A client has no possible way to know what the "best" choice from a list
of RR DNS IP's might be, whereas server products like F5's 3DNS/global
load balancer monitor server load/availability and and an assortment of
other things to offer the best choice. I don't know if any of these
products send an ordered list in the preferred choice, but things would
be more robust if they did and the client walked the list in the
presented order. If the server must send only one IP to control the
client's choice, it causes a problem when circumstances change during
the client application's run time. Many clients (particularly IE - the
most popular one..) will cache the list given in the initial lookup and
on subsequent failure will retry others from that list but will not do
another DNS lookup as part of the error recovery. Thus if they have
been given a preferred and backup IP in the first lookup, they'll keep
working even if the preferred server becomes unreachable. But, if they
resort the list, half of them will pick the wrong first choice even when
it is reachable.
--
Les Mikesell
lesmikesell@xxxxxxxxx
--
fedora-devel-list mailing list
fedora-devel-list@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/fedora-devel-list
