Re: HTTP router point-of-view concerns

[James M Snell <[e-mail]>]

Most of the feedback that I've seen around header compression can be
summarized as:

  Sending less data on the wire... Good.
  Less latency... Good.
  More efficiency... Good.
  Less visibility in the protocol... Bad.
  Complicated implementation... Bad.
  Introducing statefulness in the protocol... Bad.

Where I come down on this is simple:

Yes, we need a way of encoding headers more efficiently in HTTP/2
relative to HTTP/1... however, as the mantra states, Perfect is the
Enemy of Good. The header compression algorithm as currently stands is
*technically* a very good approach, but that doesn't mean it's the
right approach for right now.

The problems we see in the verbosity and repetitiveness of HTTP
request and response headers has more to do with how those headers are
defined and used that it does with how those headers are encoded on
the wire. There are reasonable application-level approaches we could
be exploring to reduce redundancy and increase efficiency that would
not require complex new mechanisms and requirements to be added to the
base framing protocol. For instance, I have demonstrated many times
that by simply changing the way we define header values, we can make
significant reductions in header transmission size.

>From everything that I can see so far, Header Compression is
completely orthogonal to the basic operation of the new HTTP/2 framing
layer. It is not critical for the new framing layer to be successful.
Yes, it has it's benefits. Yes, it makes things more efficient. Yes,
it has security advantages relative to other compression options. But
no, it is not the Simplest Thing That Could Possibly Work and no, it's
not technically required in order to make HTTP/2 successful... and so
far, the majority public reaction seems to be "Whoa! WTF is with all
the new complexity! Dial it back a bit please!"

It's good that people are starting to get implementation experience
with what's been written up so far. It's good to start getting interop
testing going. For me, going through my own implementation of current
header compression mechanism has simply reinforced what I had already
suspected and feared: it's a lot of very well designed and functional
new complexity that has some benefit but is not strictly required to
get the job done.. and, what's worse, is the fact that it makes us
jump through some additional design hoops (e.g. the routing question)
that would otherwise be fairly simple to address.

That all said, I'm definitely willing to be convinced otherwise. Let's
see what happens in Germany next month, but let's definitely keep an
open mind towards alternative, less complicated (albeit, possibly less
efficient) approaches that do not incur such a large New Complexity
Tax.

- James


On Fri, Jul 12, 2013 at 9:36 AM, Roberto Peon <[e-mail]> wrote:
> This was the first thing I experimented. :)
> It either requires two different state size settings, or it makes state size
> management .... interesting....
> Having a single table made much more sense and was less complicated,
> especially for proxies.
>
> -=R
>
>
> On Fri, Jul 12, 2013 at 8:22 AM, Jeff Pinner <[e-mail]> wrote:
>>
>>
>> On Fri, Jul 12, 2013 at 2:11 AM, Mike Belshe <[e-mail]> wrote:
>>>
>>> I'm also in favor of removing the compressor completely.
>>
>>
>> So the compressor buys us the ability to share headers between streams and
>> possibly to reduce the size of the headers via some sort of encoding
>> (whether it's typed encodings, or huffman compressed strings, or varint
>> lengths, etc). So a dumb proposal:
>>
>> A HEADERS frame consists of encoded name values pairs, let's say varint
>> length followed by UTF-8 bytes of the string (we can argue over compressed
>> strings, types, etc. later, but basically no indexing into shared state).
>>
>> Sending a HEADERS frame on Stream-ID 0 creates a set of headers that gets
>> saved and added to the HEADERS frame that opens any streams after it is
>> sent. Sending a new HEADERS frame on Stream-ID 0 overwrites the previous
>> frame.
>>
>> This allows us to share Cookies, User-Agent, Host, etc. between requests,
>> but wouldn't allow for any response header sharing. It would allow us to
>> share headers for pushed responses since those are streams opened by the
>> server.
>>
>>
>