Re: HTTP/2 Priorities Proposal

I like it that this solves the proxy problem in its entirety.
Since I do proxy nearly everything, as do many large sites and CDNs, this
seems like a good thing.

-=R


On Tue, Apr 8, 2014 at 5:50 PM, Tatsuhiro Tsujikawa
<tatsuhiro.t@gmail.com>wrote:

>
> 2014/04/09 8:26 "Jeff Pinner" <jpinner@twitter.com>:
>
> >
> > I believe it would be simpler (both conceptually and in practice) to
> slightly modify the priority structure.
> >
> > Currently we have group-ids with weights and (possibly-exclusive) stream
> dependencies. The default is for the a stream to create a new group with an
> id equal to its stream identifier and some default priority.
> >
> > Conceptually this is the same as saying that the priority is a
> non-exclusive stream dependency of 0, but this dependency is special in
> that it is weighted.
> >
> > When depending on an existing stream, the stream has an implicit weight
> of 1 / total # of dependent streams (rounded appropriately to integers).
> >
> > This conceptually creates a "weighted groups of dependency trees" but it
> seems that this can equivalently be thought of as a single "weighted tree"
> with implicit or explicit weights.
> >
> > I am suggesting that it would be easier both conceptually and in
> implementation to replace the idea of groups entirely with weighted
> dependencies. Implementations already have to handle weights at some point
> to deal with groups. Thus determining priority is the traversal of a single
> weighted tree.
> >
> > The on-the-wire implementation then only requires a single flag for
> signaling priority information, with the frame layout always containing a
> weight and a stream dependency (possibly 0) with an "exclusive" bit.
> >
> > The defaulted priority information is then a dependency of 0 with the
> default weight (16). This is identical to a new group with the group-id ==
> stream-id.
> >
> >
> > I believe this proposal has the benefit in that the wire format is
> simplified, the overall complexity of the implementation is reduced, and
> all priority information is proxy-able. The drawback is that in the
> existing implementation, you can add a stream to a named group and share
> it's weight is automatically divided equally amongst the roots of the trees
> in the group. In this implementation, doing so would require sending
> priority frames to re-weight the root streams.
> >
> >
> > Thoughts?
> >
>
> The one dependency tree is easily collapsed to multiple trees in usual
> situation. For example, if 5 streams depend on stream A, and A is closed,
> those 5 streams become root for each subtree. And changing weight of them
> requires 5 priority frames. If the benefit of this proposal outweighs this
> cost, it is worth considering. I don't know proxy case outweighs this
> though. If changing weight for a group is happening frequently and it is
> the base idea of dependency based priority, then I think the keeping -11 is
> way to go.
>
> As for framing, the complexity of current format is very small compared to
> the rest of priority handling.
>
> Best regards,
> Tatsuhiro Tsujikawa
>