ALPN negotiation (was Re: Add extension work to Interop matrix)

[Christian Huitema <[e-mail]>]

On 1/7/2020 4:31 AM, Dmitri Tikhonov wrote:
> Hello,
>
> There are now a few extensions that have been proposed and implemented.
> We should add them to the next Interop matrix.  I suggest using the
> following codes:
>
>     g   Datagram extension
>     r   HTTP priorities extension
>     q   Loss bits extension ('l' looks too much like '1')
>     1   One-way delay extension
>
> 'g' and 'r' can be added to the third row, while 'q' and '1' can be
> added to the second row (dropping the "use uppercase letters only"
> requirement).

Looking at https://interop.seemann.io/, it seems that the worst interop
problem so far is the H09/H3 split. Of the 7 implementations listed
there, we see 2 that do not support H3 and at least one that does not
support H09.

Most of the failures in the "measurement" matrix are related: clients
and servers are programmed to request or serve only H09, and fail if the
peer does not support it. But some propose a list of supported ALPN and
the ALPN negotiation succeeds.

It seems that ALPN negotiation is going to be a practical requirement
going forward. Not so much for negotiating H09 versus H3, we can expect
H09 to fade away at some point. But we will have to negotiate h3-24 vs
h3-25 and similar transitions for a good bit of time, and then we will
probably move to h4-00, h4-01, etc. So maybe we should start testing that.

-- Christian Huitema