Crossref has decided to change the HTTP redirect code used by our DOIs from
303 back to the more commonly used
302. Our implementation of 303 redirects back in 2010 was based on recommended best practice for supporting linked data identifiers. Unfortunately, very few other parties have adopted this practice.
At this point, we need to emphasise that we have never seen our use of 303s actually affect page rankings. But at the same time, use of 303 redirects has not had wider uptake. Maintaining this quixotic behaviour just isn’t worth the effort. We hope that, in the future, we can use other techniques (e.g. signposting & cite-as) to achieve some of the things that 303 was supposed to do.
Note that these changes will not affect users or machines using DOIs. The change should be entirely transparent.
Below we provide some background to our decision and after that we provide some detailed technical notes from Jonathan Rees and Henry Thompson who have been very kind in helping to provide Crossref technical guidance on how we can help DOIs best support linked open data and adhere to HTTP best practice.
Back in 2010, Crossref, DataCite (and later, several other RAs) responded to concerns that DOIs were not “linked-data friendly.” There were three problems with DOIs at that time:
On the advice of several people in the linked data community, we proposed some options for fixing this. And we finally settled on:
By any practical measure, machine use of DOIs has exploded since we made these decisions back in 2010. Crossref’s APIs and content negotiation handle over 800 million requests for machine readable data a month. Our sibling organisation, DataCite, has also seen a huge growth in machine use of DOIs. Many applications, from bibliographic management tools, to authoring systems and CRIS systems, make use of machine actionable DOIs all the time. So clearly our work to promote DOIs as machine actionable identifiers is working, but we are certain that our current use of 303 redirects has nothing to do with this growth.
First of all, as we said, very few parties have actually subscribed to the notion of using 303s to help distinguish “the thing” from “something about the thing”.
Secondly, even if they did try to rely on 303s to make this distinction, they would quickly get confused because the DOI is so often just the first in a chain of redirects which do not implement the same semantic distinction. At this point we should be clear - Crossref thinks these kinds of long redirect chains are a bad idea for two main reasons:
But we also cannot legislate them away. They exist. And in the real world you will find plenty of DOIs that do a 303 redirect to a system that, in turn, does a 302 redirect to a system that does a 301 redirect and…eventually ends up someplace returning a 200. You get the picture. How on earth is a machine supposed to interpret a 303->302->301->302 redirect chain?
Furthermore - nowadays, after following this chain of redirects, you will often find yourself on a “page” that is both a landing page and the article itself. Dynamic, one-page applications can simply morph the one into the other without the use of additional HTTP requests.
In other words, using 303s is not helping machines interpret what the DOI is pointing at. And yet, people seem to be making good use of machine actionable DOIs and they are not complaining much about it.
Personally, I’d might have just been happy to switch back to using 302s simply so that I could cut down on my conversations with SEO hacks. But that wouldn’t be a principled approach. In 2010 we spent a lot of time considering the initial switch to 303s- we needed to consult with the LOD community on a potential switch back to 302s. At the January 2018 PIDapalooza I had a chance to talk to Henry Thomson about the 302/303 dilemma we faced, and he along with Jonathan Rees very generously provided the following feedback.
If one goes to the trouble to organize an identifier system, then the desire that such a system should last as long as possible leads one to aspirationally say it’s a persistent identifier (PID) system. The unwillingness of the major browser suppliers to implement new URI schemes for PIDs initially hindered their use on the Web and this in turn inhibited widespread adoption. More recently a number of PID approaches have enjoyed very rapid growth as a result of a compromise: these PIDs participate in the World Wide Web by defining simple conversion rules mapping identifiers to actionable (‘http:’ and/or ‘https:‘) forms and providing resolution servers that redirect requests for such forms to the appropriate destination.This approach has been widely adopted and is very successful, because it is so useful. An identifier’s actionable form leads, via the HTTP protocol and one or more redirections, to a web page that bears on the ground identity of the associated entity – or perhaps even directly to the entity itself, if the system is one for document entities that are naturally provided as web pages. The nature of the retrieved web page varies from one system to the next.
A confusion arose, however, over claims in various technical specifications (URIs, HTTP, Web Architecture) that the normal case is for the protocol to yield a “representation” of the “resource” “identified” by the URI. None of these terms is adequately defined by the specifications, and initially the language was not taken as normative. Those deploying identifier systems took the HTTP “resource” to be the entity associated with an identifier, and understood the “resource” as being “identified” by the URI, but it was never clear what was, or wasn’t, a “representation” of a given entity/resource: a description of the resource, the resource itself, a version of the resource, instructions on how to find the resource, etc. Sixteen years ago, in an attempt to clarify the intent of this part of the theory of URIs, and to allow applications to usefully and uniformly exploit the idea that an HTTP 200 response must deliver a “representation” of the “resource”, Tim Berners-Lee asked the W3C Technical Architecture Group to consider what came to be known as the httpRange-14 issue. It’s now 13 years after the TAG gave advice which almost no one was happy with, and 5 years after work on issue httpRedirections-57 (which superseded httpRange-14) ground to a halt. There’s still no consensus on whether it’s OK to return landing pages with a 200 status in response to requests for pictures or publications, but the Web seems to be working nonetheless, and no one seems to be bothered much anymore.
The provision of HTTP-based resolution services has stimulated widespread support for the use of identifier systems with Web resolution, particularly in the scholarly journal publication context. Those setting up HTTP resolvers responsible for identifier systems must decide which HTTP response code should be used. The TAG’s advice sows doubt on the use of the 200 response code when the response would have been a landing page, and many resolvers avoid 200 regardless and use redirection for administrative purposes, for example
‘10.1109/5.771073’, or ‘https://identifiers.org/uniprot/A0A022YWF9’ to
‘http://www.uniprot.org/uniprot/A0A022YWF9’ for the Uniprot identifier
So the response should be a redirection, but what kind, 301, 302, or 303? (Or 307, which is almost the same as 302.) A 301 redirect seems to say that the URI is not persistent (since its target is deemed “more persistent”). A 302 redirect seems to say that the response could have come via a 200, and so suffers the same fate as 200. That leaves 303, as hinted at in the TAG’s advice. This idea got some traction: Ten years ago a Semantic Web interest group promoted the TAG’s advice in a published note, and seven years ago one of us wrote a blog post giving the same advice for resolvers for PIDs in publishing.
However, not only is there neither consensus nor general utility around this strict understanding of the use of the various response codes – that is, that resolution to a landing page is inconsistent with a 200 (and a posteriori therefore with a 302) – but also the range of usage patterns for redirection of HTTP requests has grown and ramified over time as the Web has grown and become more complex. It’s on the face of it unlikely that a mere three response codes can capture all the resulting complexity or cover the space of outcomes (in terms of e.g. what ends up in the browser address bar or what search engines index a page under) that a page owner might like to signal.
We find in practice that some PID redirections are ending up (usually after further publisher-local redirects) at the “identified” document, some at landing pages, and some at one or the other depending on the requesting site, for example in the case of paywalled material.
In the absence of a rethinking of the whole 3xx space, it seems to us that only the 301 vs. 302 distinct ion (roughly, 301 = permanent = please fix the link, and 302 = temporary = don’t change the link) is well understood and more or less consistently treated, whereas for 303, web servers are not very consistent and both search engine and citation crawler behaviours are at best inconsistent and at worst downright unhelpful.
So, we believe it is in both users’ and publishers’ interests for resolvers of actionable-form PIDs to use 302 redirects, not 303.
If we want to help machines better understand the resource that a DOI points at, we have to explore using more nuanced mechanisms.
Just using 302 for the first redirect doesn’t do everything necessary to effectively support the emerging PID+redirection architecture. It’s at the end of the redirect chains that we need more: a standardised way to find the PID back at the start of the chain. The ‘cite-as’ proposal does exactly this, and we hope it’s quickly approved and widely adopted. Once that happens a proposal for augmenting browser (and API) behaviour to prefer, or at least offer, the ‘cite-as’ link for bookmarking and copying will be needed.
2019 November 09
2019 November 09
2019 October 29