On Monday, December 21, 2015 01:13:10 PM Tony Finch wrote: > Paul Vixie <p...@redbarn.org> wrote: > > any HTTP initiator who wants out of order response processing will have > > to negotiate for it (see mogul's 2001 RID draft) and will then have new > > responsibilities for matching up the out of order HTTP responses with > > then-outstanding HTTP requests. > > The current way to deal with out-of-order responses and head-of-line > blocking in HTTP is HTTP/2.
since http/2 is a completely new protocol, i think that's a strange way to say it. as you point out below, existing HTTP initiators accomplish out-of-order processing by using multiple parallel HTTP/1.1 TCP sessions. how the HTTP asynchrony is achieved is beyond the scope of an HTTP application spec such as DNS-over-HTTP. i will suggest some text below. > > If you do DNS over HTTP then there has to be an exact correspondence > between HTTP requests and responses and DNS requests and responses - > anything else would be madness. This implies that DNS over HTTP/1 only > supports in-order pipelined queries and responses in each connection; to > avoid head-of-line blocking you need either multiple connections or > HTTP/2. by all means let us add text to the effect that "HTTP/1.1 or later can be used, with the DNS- over-HTTP initiator having the responsibility to discover what HTTP protocol version the responder supports, and to speak that protocol version correctly -- for example, if an HTTP protocol version supports multiple outstanding requests and out of order responses, then the DNS-over-HTTP initiator is required to match inbound HTTP responses with outstanding DNS requests." -- P Vixie
_______________________________________________ DNSOP mailing list DNSOP@ietf.org https://www.ietf.org/mailman/listinfo/dnsop
