On Wed, Nov 04, 2020 at 03:07:22PM +0100, Emeric Brun wrote:
> Hi Dinko,
> 
> > Sadly I haven't had Kube-DNS anywhere and i think that CoreDNS is supposed 
> > to be way to go from Kube-DNS. Hope this helps.
> It does.
> 
> Really appreciate!

Actually this tends to confirm our preliminary thoughts, which are that
if some people absolutely want to load-balance DNS on haproxy, they'll
have to comply with some supported features, or look for something else.

I think it's a reasonable tradeoff because those who insist on this are
also those who want to use so-called "modern" tools (placing "modern"
and DNS in the same sentence always leaves me a strange feeling that
something 37 years old is still modern).

@Lukas, to respond to your concern, I don't like DNS either, but I much
stronger hate the noise that some people regularly make about "UDP support"
which when you ask is for syslog and dns (both being the wrong way to
proceed, proving they don't even understand what they're asking for).

Now with 2.3, syslog is handled like a proxy should do, which is not
by blindly forwarding datagrams and causing trouble on the server, but
instead by passing the messages, possibly re-encoding them according
to the desired RFC level and adapting udp/tcp/stdout/whatever in any
sense. The other one that remains now is DNS. I personally consider
that it's stupid to want to LB inbound DNS with no added value, LVS
will do it better and will scale better. In addition, if the reason is
a matter of load, then one should really wonder if the LB will not reach
limits before the server, having more than twice the work to do. However
there could be a few compelling cases where doing it correctly could
add some value. By correctly I mean that we'd typically receive mostly
UDP datagrams and aggregate them over TCP connections. In this case it
could make sense as a protocol accelerator/offloader/translator like
haproxy already does for HTTP or TLS. In addition it might be possible
to implement some short time caching to save on the internal bandwidth.
Maybe later some of the DNSSEC job could be offloaded by haproxy, I
don't know.

So that's where I draw the frontier. Blind UDP, no, only morons want
this in a userland proxy. Value-added UDP-based services, why not if
it can be demonstrated that it does really bring some value and aligns
well with what we already do around.

Cheers,
Willy

Reply via email to