- From: Alex Russell <slightlyoff@google.com>
- Date: Wed, 5 Feb 2014 18:16:48 -0800
- To: Jeni Tennison <jeni@jenitennison.com>
- Cc: "www-tag@w3.org List" <www-tag@w3.org>
- Message-ID: <CANr5HFUdymz3u3Wi6FZBcG2P9HUgytocEOZkvk8yja3dxhgX2w@mail.gmail.com>
On Wed, Feb 5, 2014 at 2:46 AM, Jeni Tennison <jeni@jenitennison.com> wrote: > Alex, > > > > Yes, it’s not as conceptually elegant a solution as I would > like either. > > > The more elegant approaches that I looked at had larger drawbacks, as > I’ve > > > documented. Can you think of another way of approaching this that I’ve > > > missed? > > > > Yehdua's proposal (changing URL parsing) avoids the issue by > > conflating addressing with package location. > > Yes, but as I’ve written [1], it has larger architectural drawbacks in > baking new syntax into URLs. Do you disagree with that analysis? If we can't overcome the per-parser issues, then I think the balance swings to changing URL parsing in a way that seeks to minimize unintended consequences. > > > I’m not convinced that having to make a request before making another > > > request for a package is a fundamental problem. I know that the point > of > > > the packages is to minimise the number of requests that are made, but > the > > > BBC home page, for example, involves 110 requests; doing 2 rather than > 110 > > > seems like a win. > > > > Again, please see the discussion of pre-parse scanners. It's unclear > > (without some other signal, say a globbing pattern that's part of the > > to indicate what resources are in the package and which aren't) that > > we will actually avoid the extra fetches in practice under the proposal. > > Sorry, I’d missed that because my email client had "helpfully" hidden it > as quoted content. Sigh. > > Does having a simple textual manifest file that just lists the content of > the package as the first item in the package help at all? > Unclear. It's a foot-race, after all. A client will do request to get the document, which usually streams in pretty quickly relative to the RTT's you need to navigate for sub-resources. This is confounded by advice like this which indicates to authors that they should be requesting some resources as early as possible in their documents: https://developers.google.com/speed/docs/best-practices/rtt#PutStylesBeforeScripts This is especially important for resources that are going to be fetched off-origin and for which warming up DNS can add considerable latency. If we consider the MTU that causes the first couple of packets to be several K of content, and an RTT for the pack file on the order of 100ms (conservative, but not outlandish, particularly on mobile), I think we'd lose the race for many resources in the <head> of a document. Not where we want to be, and it gets worse the higher the latency on the network. > Or, as you suggest, a globbing pattern, which could be a parameter on the > content type both on the link and on the response to the request for the > package. For example: > > <link href="assets.pack" rel="package" > type="multipart/package;glob=/assets/*,/favicon.png"> > > Thoughts? > I'm going to try to come up with alternatives, but I think the globbing pattern + a manifest-as-first-item strategy is the only thing that can save this from the fire. Without some knowledge of which pre-flighted requests to delay on package files, I fear pre-scanning is fatal. > Cheers, > > Jeni > > [1] https://github.com/w3ctag/packaging-on-the-web#specialising-urls > -- > Jeni Tennison > http://www.jenitennison.com/ >
Received on Thursday, 6 February 2014 02:17:46 UTC