Re: update: http progress notification

On May 27, 2010, at 4:40 PM, Adrien de Croy wrote:
> On 28/05/2010 10:45 a.m., Roy T. Fielding wrote:
>> On May 27, 2010, at 2:56 PM, Adrien de Croy wrote:
>>> OK, so HTTP/1.0 is covered in any case, since 1xx responses are not allowed to HTTP/1.0 clients.
>>> That just leaves the issue of whether a client should indicate support (or desire) for progress notifications or not.
>>> I'm a little bit reluctant to suggest the progress notifications should not require any indication from the client.  If only because some clients may want to be able to suppress generation of such messages for whatever reason.
>>> So I think my preferred option is a Progress token in the Connection header.  That way I don't have to try and dream up something useful in terms of parameters for a potential Progress request header.  It also means clients that aren't written (e.g. existing clients) to process these messages won't receive them for discarding and waste bandwidth (esp over slower high-cost circuits such as GPRS / G3 etc).
>> Sending an additional 10 to 30 bytes on every single request,
>> in the critical path, just in case the browser might access
>> such a resource (a one in a billion chance, at best), is an
>> extremely poor trade-off.
> actually if you're behind a corporate firewall that does this, it's closer to a 100% chance depending on scanning rules for the firewall AV.  When we deploy this with WinGate and AV, that will then be millions of browsers (and other HTTP agents) seeing these 103s for pretty much all requests if we don't negotiate.  I've tested Chrome, FF and IE8, but I haven't tested things like windows update, BITS service etc etc etc.  There's no way to guarantee what won't break.

Again, that is a very tiny percentage of agents when compared to the
rest of the Internet traffic and needs to be crossed with an even
smaller percentage of resources that would benefit from a progress feature.

I see no reason for such a small instance of potential benefit to
impose a constant cost.  Services like Windows Update can be easily
bypassed (why you would want to scan such a service is beyond me,
since it often delivers virus signatures inside virus-checking
client tools).

> I was of the impression switching behaviour based on User-Agent was frowned upon.  It creates at best a maintenance issue for admins (keeping the list of UAs current).

You don't have to keep a list of current agents.  You have to keep
a list of broken agents that do not properly handle HTTP/1.1 status
codes, and you can make that configurable.  Apache does that kind of
thing all the time.  It might even be the empty list, for all we know.

> I agree it will be a long time before origin servers generate them.  So probably it makes no sense for now for a proxy to forward the Progress token upstream, unless it's talking to another proxy.  So actually I don't see many of these going over the wire onto the net. The main benefit is for a client talking to a local proxy.  RFC2616-compliant proxies that don't understand Progress will already strip it.

So you would only have UAs send it when configured to use a proxy?

> I don't know of any browser that doesn't already include a Connection header in every request already for keep-alives (even through default semantics for 1.1 is to keep alive so it's largely redundant), so overhead in practise will only be an extra 10 bytes.
> Also, clients could make some choices about likelihood of benefit before adding Progress to the Connection header.  E.g. for retrieving images, JS, CSS etc it could be turned off.

Anything that relies on the client guessing the nature of a resource
is not a good plan.  Images, for example, are often generated, converted,
or cropped on the fly.

> Or behaviour could differ for proxy connections vs direct and be configurable in the browser.
> Otherwise if we don't advertise support, and it's on by default, we need to resort to things like administrative settings, e.g. admin can turn it off/on per client IP or UA or Content-Type or something.  That loads up burden on the admin.

The number of scan-proxy-maintaining admins or long-resource-processing
owners is far smaller than the groups effected by the other options.
The cost should be placed on those benefitting most from the feature.

> As for Henrik's point about benefit for non-supporting clients, where these messages will stop connection timeouts.  I can understand a client may consider something is happening, but if there is no progress indicated to the user, we are back at square one (where we are now), where the user hits retry a few times then calls his tech support or admin.  So I don't think the benefit will be there for browsers.  Other agents certainly may benefit.

Or you could just send a 202 response if the software anticipates
a lengthy delay.


Received on Friday, 28 May 2010 00:21:35 UTC