Web services are becoming more and more popular in scientific computing. There are already lots of Web sites where you can upload some data, push a button to run a computation, and download the results. Sometimes a computational Web service is integrated with a database, so you don’t even have to upload any data. You type in some search, push a button, and get your data. The computation is done somewhere in the cloud.
Web services are becoming popular because of their convenience. For the user, there is no software to install. For the software developer, there is no distribution or packaging, and no more trouble with users refusing to upgrade to the latest version. Moreover, convenience for users means more users and thus more citations for the software. On the development side, there is a price to pay, of course, because the Web layer needs to be developed in addition to the computational software. But a good desktop user interface is a lot of work as well, so in the end the Web service may well be the easier option.
So in the end, everybody wins, which is why Web services are becoming so popular. And they all lived happily ever after…
Well, not quite. There is one big problem with Web services for scientific computing: in terms of reproducibility, they are a disaster. The statement “the results in table 1 were obtained from the Web service at http://…” in a scientific publication is almost as informative as “table 1 was dictated by the Oracle of Delphi”. Well, almost, there are two important differences. First, a reference to the Oracle of Delphi would lead to certain rejection of the paper by any serious journal, whereas a reference to a Web service is considered normal. Second, Web services usually do come with a citable publication that explains what they do and provide a summary of the algorithms used. Still, Web services are the ultimate black boxes of scientific computing: you get a result but neither a version number nor an archivable copy of the current source code. A colleague trying to reproduce your work a month later may well get different results from the same input, without any explanation.
However, there is absolutely no technical reason why computations through Web services should not be reproducible. They are not because their clients don’t ask for reproducibility. As soon as scientific journals start rejecting submissions that are not fully reproducible, we will have Web services providing reproducible results. Trust me, it’s Pythia herself who told me.
After all, a Web service is just a Web-based user interface to computational software. If software on your personal computer or on your computing cluster can deliver reproducible results, so can a Web service. Nothing prevents Web service providers from putting their source code on Github and the commit number for the exact version that was used into every result file offered to the user for download.
That’s still not quite the end of the story. Imagine a near future with lots of computational Web services, and users (or scripts they write, or meta-services) feeding the results from one Web service into the next Web service. This leads to exactly the same problem of workflow logging and provenance tracking that we have today when using a chain of traditional software on a desktop computer. Users or their workflow management software will have to figure out which result was produced from which input using which Web service, and figure out everyone’s version numbers. Wouldn’t life be easier if this happened automatically? If every file passed around between Web services contained a complete trace of its origins, including references to the software that was used?
That’s where ActivePapers enters the stage. Imagine the following scenario. You prepare your input data in Excel 2020 and select “publish on figshare in ActivePapers format” from the menu. A DOI pops up, which you paste into your first Web service. The Web service does its computation on the data, using its computational software that is also published as an ActivePaper. What it sends back to you (or directly to your figshare account) is an ActivePaper as well, with references to your input data and to the software that was used. Ready to be fed into the next Web service. Provenance tracking is integrated into the workflow itself, and stored along with the data.
Now, finally, everybody wins. Scientists run computations without installing software and without worrying about data management. Software developers do exactly what they do now: develop computational software plus Web interfaces. They don’t have to worry about data management either because the decentralized ActivePapers infrastructure takes care of that. And science was reproducible ever after.