Re: draft-wullink-restful-epp-00.txt

See my earlier email Klaus but there's no reason why you can't terminate a
connection based on a policy for permitted packet flows - so I still don't
see this as any kind of a significant problem.  You do have to use a
persistent connection algorithm when using a load-balancer, however with
packet-shapers you can set policies on the maximum flow rate per connection
or even across a group of connections.  In short, you are expecting the
client to take some responsibility for not hammering the server within a
single connection and limiting them if they do.  I don't see valid use case
where a single EPP connection operating against today's equipment and
environs would need to apply an excessive amount of load for a server in the
course of any normal registry business. If you know of one, I'd love to know
about it.

Namedrop behaviours of course are another special use case, but segmented
special connection pools and rate-limiting have typically solved that one as
well.  The scope of the namedrop server hammering was also reduced by the
now active secondary marketplace that occurs prior to a name being dropped
in the first place.

Best Regards,

Michael Young

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf
Of Klaus Malorny
Sent: April-23-12 10:20 AM
To: Maarten Wullink
Cc: [email protected]; [email protected]
Subject: Re: [provreg] draft-wullink-restful-epp-00.txt

On 23/04/12 16:02, Maarten Wullink wrote:
> Stateless EPP makes it possible to load balance individual EPP 
> requests instead of EPP connections. This should make it easier to
distribute to load evenly across the servers.
> Distributing tcp connections across EPP servers does not guarantee an 
> equal load on every server. This is because not all EPP connections are
equal.
>
> Also when there is no persistent tcp connection it should be easier to 
> remove a server from the server pool for maintenance without having to 
> disconnect the tcp connections of the clients connected to the server.
>

Hi all,

while I have mixed feelings regarding the REST approach, I agree to the
arguments:

Although RFC 5734 allows the server to actively terminate the communication
with the client, it can only do this by closing the connection. There is no
graceful shutdown that would allow the client to prevent an I/O error on the
attempt to write a message. Assuming a busy connection, the sudden
interruption leaves the client in an ambiguous state whether the (partially)
transmitted request has been received and processed by the server.

For the benefit of the registrars, we avoid to close connections in our
server side EPP implementation except for those connections which are idle
over a long period. As a consequence, in a load balanced configuration, the
balancing can actually only occur via the assignment of new connections to
the server with the lowest load, so an unequal load cannot easily be
leveled.

Regards,

Klaus
_______________________________________________
provreg mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/provreg

_______________________________________________
provreg mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/provreg