Re: Proposal: Run QUIC over DTLS

[Christian Huitema <[e-mail]>]

On 3/9/2018 9:24 AM, Gabriel Montenegro wrote:

> Another interesting result of this discussion is Subodh’s suggestion to clarify the interfaces between subsystems (“layering structure”): QUIC transport and packetization vs crypto on stream 0 vs UDP. 

In fact, the current QUIC design is a good example of the evolution in
protocol architecture from "layering" to "functional decomposition". The
layering approach, as popularized by the OSI model, consider the system
as a set of filters layered on top of each other. It is a powerful
approach, but it has limitations. The biggest limitation is tied to the
design principle: each layer obscures the view of the other layers; to
mitigate that, the layer API needs a set of pass-through functions, such
as for example passing a setsockopt to TCP across TLS.

In contrast, functional decomposition looks at big functions such as
reliability, forward error correction, real time delivery, and arranges
the protocol as an assembly of these functions. This is very much the
direction pointed to by TAPS. In the case of QUIC, the relation with TLS
is an example of such functional decomposition, treating TLS as a black
box that negotiates keys.

We can certainly agree that the stream zero approach is not ideal. It
may very well be that a set of TLS oriented frame types would be much
cleaner. But in my mind a strict layering approach on top of DTLS would
not be architecturally cleaner. I believe that it would be a step backward.

-- Christian Huitema