- From: Shel Kaphan <sjk@amazon.com>
- Date: Fri, 29 Nov 1996 17:53:00 -0800 (PST)
- To: Larry Masinter <masinter@parc.xerox.com>
- Cc: sjk@amazon.com, mogul@pa.dec.com, http-wg%cuckoo.hpl.hp.com@hplb.hpl.hp.com
Larry Masinter writes: > Shel, > > Are you saying that none of the current proposals for hit metering > will help with unique-URL cache-busting because they don't actually > address the problem of browser behavior? > I am not entirely up to speed on the proposals, so I can't really comment on them, but I don't have to be that familiar with them to make the point, because what I _am_ saying is that cache-busting will continue to be practiced for reasons other than hit-metering, using techniques even more evil (from the cache abuse perspective) than pre-expiring documents. It's friendlier to caches to tell them _not_ to cache documents that are requested using unique URLs, since you know they'll never be requested again, but since pre-expiration is one of those things that makes browsers behave weirdly (within the spec, I might add), it is avoided for that reason. > # Not to beat a dead horse or anything, but the reason people use these > # techniques is because they are the only way to guarantee some degree > # of control over the user experience. > > Don't cookies do the right thing? > In their domain they do the right thing (unless they get cached in a 1.0 cache that doesn't know about cookies!) But they are not universally supported (yet), and they don't really address any of the document expiration-related problems that browsers have. Also, as long as there is any doubt about the correctness of caches, and right now there is plenty of doubt, unique URLs are _much_ more reliable than any other means to make sure that the client gets the requested document and not some out of date version, or one intended for someone else. Of course, 1.1 and future versions will improve this whole story, eventually, but if there are protocol features that remain too tricky to use because of things browsers do, the situation isn't going to improve much. Service providers care more about correctness than cache and bandwidth friendliness. > # To really beat a _thoroughly_ dead horse, this is the case because > # caches and history mechanisms are improperly conflated in most > # browsers. > > Didn't we 'fix' this with HTTP/1.1? > > Larry > > If the browser guys read the spec carefully enough, and follow the recommendations as well as the requirements, the situation will be _improved_ (eventually), but to really get this one nailed, I don't really believe that the small amount of verbiage we threw at it is going to be enough. To get it "right", we'd need to allow for the kind of browser controls in HTTP that nobody wants to contemplate because it seems to many people to be a mixing of levels that makes people uncomfortable. I can understand that concern and even sympathize with it, but that doesn't make the issues go away. If you read the spec, already understanding the issue, you'll see that the issue is addressed, but if, as an implementer, you don't already grok the architectural significance, you are unlikely to be sufficiently guided by what is in the spec. (There's only so much network specs can do). To my mind the problem is that there seems to be no forum in which it is appropriate and effective to discuss the interaction of HTTP and the subtleties of browser behavior. In this working group we just want to stick to the "bits on the wire", which is certainly understandable. But I think with HTTP we have a new kind of beast, and the problems that I'd like to be able to work on can't be solved by thinking only about bits on the wire. That said, I'm kind of resigned to the fact that this isn't going to be dealt with any time soon at the protocol level, and that, as a practical matter, one does what one has to do, which is of course very sad for the scalability of the internet, improved bandwidth usage, and flexibility of new kinds of web-based services. People will find answers but they may not be very pretty ones. --Shel
Received on Friday, 29 November 1996 18:09:54 UTC