Re: Submitted new I-D: Cache Digests for HTTP/2

[Patrick McManus <[e-mail]>]

Sometimes you will get race conditions - especially when bootstrapping
whether things are h2 or not, that result in connections that don't carry
any (or perhaps 1) transaction.. and then coalescing moves everything to
one connection going forward.. typically most of that information is cached.

When I load your page in firefox I see 46 streams ( a mix of push and pull)
associated with one connection, 0 streams associated with another, and 1
stream associated with api1.shimmercat.com.. that last one is considered an
anonymous channel and won't be coalesced into your non-anonymous mix...
firefox would in general cache that coalescing information for the rest of
the session (subject to whatever memory cached eviction logic goes on).

if that's not what you're seeing - I'm happy to work with you directly.

(related to the on topic query of cache digests.. they are per origin and
coalescing is cross origin (and as we see often late bound), so I think
sending them as multiple blobs makes the most sense.)


On Mon, Apr 11, 2016 at 11:32 AM, Alcides Viamontes E <[e-mail]>
wrote:

> Thanks Mike,
>
> Can you point to any public-facing website  where connection coalescing
> is working?
>
> We haven't been lucky at configuring connection coalescing in our website
> and that have led me to believe (perhaps mistakenly) that browsers don't
> implement it yet. This page (
> https://www.shimmercat.com/en/blog/articles/data-density/ )  for example
> pulls resources from three subdomains on the same wildcard certificate,
> same web server. Chrome creates a separate connection* for each of them
> (seen by looking to chrome://net-internals, also confirmed by looking to
> the logs of ShimmerCat). As a side-note, we use SANs at development, and
> there there is no coalescing either. Same situation with Firefox.
>
> * Actually Chrome opens  an entire connection pool for
> media.shimmercat.com with six sockets, of which only one is used.
>
> On Mon, Apr 11, 2016 at 4:57 PM, Mike Bishop <[e-mail]
> > wrote:
>
>> I believe that Chrome and Firefox coalesce when DNS/SANs permit.  Edge/IE
>> currently does not.
>>
>>
>>
>> *From:* Alcides Viamontes E [mailto:[e-mail]]
>> *Sent:* Sunday, April 10, 2016 1:16 AM
>> *To:* Martin Thomson <[e-mail]>
>> *Cc:* Kazuho Oku <[e-mail]>; HTTP Working Group <
>> [e-mail]>
>> *Subject:* Re: Submitted new I-D: Cache Digests for HTTP/2
>>
>>
>>
>> Hello Martin,
>>
>>
>>
>> We will take a closer look at this. But what is the state of coalescing
>> regarding browser implementations?
>>
>>
>>
>> On Fri, Apr 8, 2016 at 3:53 PM, Martin Thomson <[e-mail]>
>> wrote:
>>
>> On 8 April 2016 at 10:44, Kazuho Oku <[e-mail]> wrote:
>> > So the question boils down to two.
>> >
>> > Q1. How many fresh resources exist per origin?
>>
>> > Q2. What is the appropriate value for P?
>>
>> I have some others:
>>
>> What is the effect of connection coalescing on this? When coalescing,
>> Is a client better off sending a separate digest for each origin, or
>> is a single large blob better?
>>
>>
>>
>
>