Re: [DNSOP] Should we try to work on DNS over HTTP in dnsop?

[Paul Vixie <[e-mail]>]

On Friday, December 18, 2015 05:56:19 PM Mark Delany wrote:
> > > That the request pipeline order doesn't necesarily match the response
> > > pipeline order is particularly unexpected in some protocols (and
> > > likely non-compliant), such as HTTP < 2.0
> > 
> On 17Dec15, Paul Vixie allegedly wrote:
> > i think this is opaque to the dns-over-http specification. that is, while
> > http may under its covers do all kinds of asynchronous things
> 
> I may not be being very clear or I may misunderstand. As far as
> implementation is concerned you are of course correct. As far as the
> semantics of the order of responses over HTTP, not so much.
> 
> What I was getting at is that a dns-over-http client can pipeline send
> request id=1 then request id=2 and get back pipelined response id=2
> then response id=1 over a single connection:
> 
> Client                    Server
>  Request id=1  ---->
>  Request id=2  ---->
> 
> .. time passes
> 
>              <---- Response id=2
>              <---- Response id=1
> 
> That is, a dns-over-http client has to do more elaborate
> request/response matching with id (+query RR) rather than rely on
> submit order for matching.

http reordering would be handled at the http layer. for example, in libcurl. if the http protocol 
agent decides to transmit multiple outstanding queries on a single tcp socket, then it will use 
its own methods to match up the eventual responses to the then-outstanding requests. 

while the following internet draft is expired, it explains these issues well, including the 
background, and the relationship to persistent connections:

https://tools.ietf.org/html/draft-mogul-http-ooo-00

however, i think that while pipelining is standardized...

https://en.wikipedia.org/wiki/HTTP_pipelining

...the responses will be in order, unless you open multiple TCP sessions.

> This re-ordering of responses is not allowed in regular HTTP 1.* and
> would almost certainly break common HTTP clients that expect pipeline
> order to be maintained.
> 
> I'm not saying this is good or bad, just saying it's an important
> topic for any dns-over-*stream* protocol to discuss. Least surprise
> and all that.

any discussion of out-of-order HTTP responses would have to explain how to match up 
responses with requests, as mogul's expired draft does above. since DNS-over-HTTP does not 
call for out-of-order HTTP responses, it does not have to explain how to match up responses 
with requests.

-- 
P. Vixie