Re: New Version Notification for draft-vkrasnov-h2-compression-dictionaries-01.txt

[Martin Thomson <[e-mail]>]

On 4 November 2016 at 01:45, Patrick McManus <[e-mail]> wrote:
> We should also be congnizant of the recent movement towards
> "content-encodings should be self describing" that came out of the encrypted
> object draft discussion - which sdch doesn't quite match (yet?). Perhaps a
> brotli with a uri for its custom dictionary combined with push of the dict
> gets us there (handwave).


I think that the main innovation in Vlad's proposal is that he doesn't
necessarily rely on a separate dictionary resource, he uses real
resources to prime the dictionary.  That means that you never spend
bytes on stuff that isn't immediately useful to you.  I'd be very
interested in seeing what effect that has on things like
responsiveness.

I agree with you that the content encoding being self-descriptive is
valuable, but I'm not quite seeing how we can get both of these things
just yet.  Maybe URLs are the wrong answer and we simply need to
identify a shared context somehow.

If you take Jyrki's observation that you might want multiple
dictionaries, you can have a header that includes a list of
dictionaries you want to use, plus optionally the dictionary that you
want to append to with this response.

Spitballing here... would identifying the context with a hash of its
state work?  Using hashes means that you don't ever have a problem
with state synchronization.  Synchronization is very much a problem if
you move to something like QUIC where inter-stream ordering is loose,
and it acts as a sanity check that you aren't corrupting data.