We have renewed our partnership with DOAJ to focus on a new set of objectives that reflect both organisations’ commitment to improving sustainable and equitable services and infrastructure. This renewed collaboration focuses on improving the quality of scholarly metadata while expanding support for journals in low- and middle income- countries.
We have worked together since 2021, primarily to encourage the dissemination and use of scholarly research using online technologies, and regional and international networks, partners and communities. This partnership has helped to build local institutional capacity and sustainability within the global scholarly communication ecosystem. A continued partnership also reflects that we have a shared community; currently almost 90% of DOAJ journals are represented in Crossref.
With key milestones achieved in 2025, including the appointment of new Directors of Technology and Programs, a move to the cloud, and some key schema updates, we now have a firm foundation for our next challenge: a redesign of our core technical systems to make them more modern, robust, and easier to maintain and scale.
In an ongoing effort to make more of our operations transparent, we have decided to start sharing summaries of our board meetings on the blog. We already post our board resolutions, but the summaries will give a bit more information on what the board discusses that may or may not show up on the list of resolutions.
When each line of code is written it is surrounded by a sea of context: who in the community this is for, what problem we’re trying to solve, what technical assumptions we’re making, what we already tried but didn’t work, how much coffee we’ve had today. All of these have an effect on the software we write.
By the time the next person looks at that code, some of that context will have evaporated. There may be helpful code comments, tests, and specifications to explain how it should behave. But they don’t explain the path not taken, and why we didn’t take it. Or those occasions where the facts changed, so we changed our mind.
Some parts of our system are as old as Crossref itself. Whilst our process still involves coffee, it’s safe to say that most of our working assumptions have changed, and for good reasons! We have to be very careful when working with our oldest code. We always consider why it was written that way, and what might have changed since. We’re always on the look out for Chesterton’s Fence!
Leaving a Trail
We’re building a new generation of systems at Crossref, and as we go we’re being deliberate about supporting the people who will maintain it.
When our oldest code was written, the software development team all worked in an office with a whiteboard or three, and the code was proprietary. Twenty years later, things are very different. The software development team is spread over 8 timezones. Thanks to POSI, all the new code we write is open source, so the next people to read that code might not even be Crossref staff.
Working increasingly asynchronously, without that whiteboard, we need to record the options, collect evidence, and peer-review them within the team.
So for the past couple of years the software team has maintained a decision register. The first decision we recorded was that we should record decisions! Since then we have recorded the significant decisions as they arise. Plus some historical ones.
These aren’t functional specifications, which describe what the system should do. It’s the decisions and trade-offs we made along the way to get to the how. Look out for another blog post about specifications.
By leaving a trail of explanations as we go, we make it easier for people to understand why code was written, and what has changed. We’re writing the story of our new systems. This makes it easier to alter the system in future in response to changes in our community, and the metadata they use.
Difficult Decisions
There are some fun challenges to building systems at Crossref. We have a lot of data. Our schema is very diverse, and has a vast amount of domain knowledge embedded in it. It’s changed over time to accommodate 20 years of scholarly publishing innovations. Our community is diverse too, from small one-person publishers with a handful of articles, through to large ones that publish millions.
What might be an obvious decision for a database table with a thousand rows doesn’t always translate to a million. When you get to a billion, things change again. An initially sensible choice might not scale. And a scalable solution might look over-engineered if we had millions of DOIs, rather than hundreds of millions.
The diversity of the data also poses challenges. A very simple feature might get complicated or expensive when it meets the heterogeneity of our metadata and membership. What might scale for journal article or grant metadata might not work for book chapters.
The big decisions need careful discussion, experimentation, and justification.
2NF or not 2NF
One such recent decision was how we structure our SQL schema for the database that powers our new ‘relationships’ REST API endpoint, currently in development.
The data model is simple: we have a table of Relationships which connect pairs of Items. And each Item can have properties (such as a type). The way to model this is straightforward, following conventional normalization rules:
We built the API around it, and all was well.
We then added a feature which lets you look up relationships based on the properties of the subject or object. For example “find citations where the subject is an article and the object is a dataset”. This design worked well in our initial testing. We loaded more data into it, and it continued to work well.
And then, the context changed. Once we tested loading a billion relationships in the database, the performance dropped. The characteristics of the data: size, shape and distribution, reached a point where the database was unable to run queries in a timely way. The PostgreSQL query planner became unpredictable and occasionally produced some quite exciting query plans (to non-technical readers: databases are neither the time nor the place for excitement).
This is a normal experience in scaling up a system. We expected that something like this would happen at some point, but you don’t know when it will happen until you try. We bounced around some ideas and came up with a couple of alternatives. Each made trade-offs around processing time, data storage and query flexibility. The best way to evaluate them was to use real data at a representative scale.
One of the options was denormalisation. This is a conventional solution to this kind of problem, but was not our first choice as it involves extra machinery to keep the data up-to-date, and more storage. It would not have been the correct solution for a smaller dataset. But we had the evidence that the other two approaches would not scale predictably.
By combining the data into one table, we can serve up API requests much more predictably, and with much better performance. This code is now running with the right performance. Technical readers note that this diagram is simplified. The real SQL schema is a little different.
Without writing this history down, and explaining what we tried, someone might misunderstand the reason for the code and try to simplify it. Decision record DR-0500 guards against that.
But one day, when the context changes, future developers will be able to come back and modify the code, because they understand why it was like that in the first place.
