Computational reproducibility has become a topic of much debate in recent years. Often that debate is fueled by misunderstandings between scientists from different disciplines, each having different needs and priorities. Moreover, the debate is often framed in terms of specific tools and techniques, in spite of the fact that tools and techniques in computing are often short-lived. In the following, I propose to approach the question from the scientists' point of view rather than from the engineering point of view. My hope is that this point of view will lead to a more constructive discussion, and ultimately to better computational reproducibility.
A few days ago, a discussion in my Twitter timeline caught my attention. It was about a very high-level model for the process of scientific research whose conclusions included the affirmation that reproducibility does not improve the convergence of the research process towards truth. The Twitter discussion set off some alarm bells for me, in particular the use of the term "reproducibility" in the abstract, without specifying which of its many interpretations and application contexts everybody referred. But that's just the Twitter discussion, let's turn to the more relevant question of what to think of the paper itself (preprint on arXiv).
A few years ago, I decided to adopt the practices of reproducible research as far as possible within the technical and social constraints I have to live with. So how reproducible is my published code over time?
In discussions about computational reproducibility (or replicability, or repeatability, according to the preference of each author), I often see the argument that reproducing computations may not be worth the investment in terms of human effort and computational resources. I think this argument misses the point of computational reproducibility.
Two currently much discussed issues in scientific computing are the sustainability of research software and the reproducibility of computer-aided research. I believe that the communities behind these two ideals should work together on taming their common enemy: software collapse. As a starting point, I propose an analysis of how the risk of collapse affects sustainability and reproducibility.
The importance of reproducibility in computer-aided research (and elsewhere) is by now widely recognized in the scientific community. Of course, a lot of work remains to be done before reproducibility can be considered the default. Doing computational research reproducibly must become easier, which requires in particular better support in computational tools. Incentives for working and publishing reproducibly must also be improved. But I believe that the Reproducible Research movement has made enough progress that it's worth considering the next step towards doing trustworthy research with the help of computers: verifiable research.
Yesterday I participated (as a visitor) in the kickoff meeting for OpenDreamKit, where one recurrent topic of discussion was notebooks, both Jupyter and Sage, including the question if they could be brought together. This reminded me of a recent blog post by Kirill Pomogajko entitled "Why I don't like Jupyter". And it reminded me of my own long-term project of integrating Jupyter with my ActivePapers system for reproducible research. That's three reasons for writing down my thoughts about notebooks and their role(s) in computational research, so here we go.
One key observation is in Gaël Varoquaux's comment on Kirill's blog post: using Jupyter for doing science creates a lock-in, because all collaborators on a project must agree on using Jupyter. There is no other tool that can be used productively for working with notebooks. It's a case of "wordization": digital content is taken hostage by a tool that defines a storage format for its own convenience without much consideration for other tools, be they competing or complementary. Wordization not only restricts the users' freedom to work with their data, but also creates headaches for the future. A data format defined by a tool can easily become unusable as the tool evolves and introduces incompatibilities, or of course if it disappears. In the case of Jupyter, its developers have always provided upgrade paths for notebooks between versions, but at some time this is bound to create trouble. Bugs are a fact of life, and I don't expect that the version-2-compatibility-feature will get much testing in Jupyter version 23. To make it worse, a Jupyter notebook can depend on third-party code that implements embedded widgets. This is one of the reasons why I don't use Jupyter for my research, although I am a big fan of using it for teaching. The other reason is that I cannot usefully link a notebook to other relevant information, such as code and data dependencies. Jupyter doesn't provide any functionality for this, and they are hard to implement externally exactly because of wordization.
Wordization is often associated with evil intentions of market dominance, as they are regularly assumed for a company like Microsoft. But I believe that the fundamental cause is the obsession with tools over content that has driven the computing industry for many years. The tool aspects of a piece of software, such as its feature list and its user interface, are immediately visible. On the contrary, its data model attracts attention only by a few specialists, if at all. Users feel the consequences of bad (or absent) data model design through the symptoms of wordization, in particular lock-in, but rarely understand where it comes from. Interestingly, this problem was also mentioned yesterday at the OpenDreamKit meeting, by Michael Kohlhase who discussed the digital representation of mathematical knowledge and the difficulty of exchanging it between different software tools. I have written earlier about another aspect, the representation of scientific models in computational science, which illustrates the extreme case of tools having absorbed scientific content to the point that its users don't even realize that something is missing.
Back to notebooks. Let's forget about tools for the moment and consider the question of what a notebook actually is, as a digital document. I think that notebooks are trying to be two different things, and that many of the problems we have with them come from this ambiguity. One role of notebooks is the documentation of computational work as a narrative with direct access to the data. This is why people publish notebooks. The other role is as a protocol of interactive explorative work, i.e. the computational scientist's equivalent of a lab notebook. The two roles are not completely unrelated, but they still significatively different.
To see the difference, look at how experimental scientists worked in the good old days of pencil, paper, and the printing press. As experiments were done, all the relevant information (preparation, results, …) was written down, immediately, with a time stamp, in the lab notebook. Like a bank ledger, a lab notebook is an immutable protocol of what happened. You don't go back and change earlier entries, that would even be considered fraud. You just add information at the end. Of course, the resulting protocol is not a good way to communicate one's findings. Therefore they are distilled and written up in a separate narrative, which surrounds a description of the work and its most important results by a motivating introduction and summarizing conclusions. This is the classic scientific article.
Today's computational notebooks are trying to be both protocol and narrative, and pretend that there is a fluent transition between them. One unfortunate consequence is that computational protocols disappear as they are edited to become narratives. This could be alleviated by keeping notebooks under version control, but I have yet to see good versioning support in any notebook-type tool. But, fundamentally, today's notebook tools don't encourage keeping a protocol. They encourage frequent changes to the code and the results, keeping only the latest version. As editors for narratives, notebook tools are also far from ideal because they encourage interactive execution of small code snippets, making it easy to lose track of what was actually executed and in what order. In Jupyter, the only way to ensure a coherent narrative is to (1) restart the kernel and (2) re-execute all cells. There is not even a single menu entry for this operation. Actually, I wonder how many Jupyter users are aware that they must restart the kernel before re-executing all the cells if they want to ensure reproducibility.
With all that said, here is my current idea of what a notebook should look like at the bit level. A notebook data model should have two distinct entries, one for a protocol and one for a narrative. The protocol entry is a sequence of code cells and results, as they were executed since the start of the computation (for Jupyter, that means the last kernel restart). The narrative is a user-edited sequence of code cells, documentation cells, and results. The actual cell contents could well be shared between the two views: store each cell with a unique ID, and make the protocol and the narrative simple lists of IDs. The representation of code and documentation cells in such a data model is straightforward, though there's a huge potential for bikeshedding in defining the details. The representation of results is much more difficult if you want to support more than plain text output. In the long run, it will be inevitable to define clear data models for every type of display widget, which is a lot of work.
From the tool point of view, the current Jupyter interface could be complemented by a non-editable protocol view. I'd also like to see a single command (menu/keyboard) for the "clean slate" operation: save the current state as a snapshot (or commit it directly to version control), restart the kernel, and re-initialize the protocol to an empty list. But what really matters to me is the data model. Contrary to the current one implemented in Jupyter, the one outlined above could be integrated into workflow management and archivation tools, such as my own ActivePapers. We'd probably see an Emacs mode for working with it as well. Plus pretty-printing tools, analysis tools, etc. We'd see an ecosystem of tools working with notebooks. A Dream of Openness.
While reading the final report of the reproducibility workshop at XSEDE14, I noticed a statement that I encounter frequently in discussions about reproducible research:
"One general consensus was that bitwise reproducibility is often an unrealistic expectation"
In the interest of clarity, let me start by pointing out that within the systematic terminology that I am trying to adopt (see this post for an explanation), I will write "bitwise replicability" from now on, as the problem falls into the technical domain (getting the same result from running the same program on the same data) rather than into the scientific one (verifying a result with similar but not identical methods and tools).
The particularity of bitwise replicability is that is almost always brushed aside as "unrealistic", which prevents any discussion about its possible importance in computational science. The main point of this post is to explain why I consider bitwise replicability important, but first of all I need to get the label "unrealistic" out of the way.
"Unrealistic" means more or less "possible in principle but impossible given various real-life contraints", and therefore the term should always be qualified by listing the constraints that make something impossible. In the context of bitwise replicability, which always refers to floating-point computations, the main constraint is that floating-point arithmetic is incompletely specified in most of today's programming languages, and that whatever specification there is is incompletely implemented in many of today's compilers. This is a valid reason for proclaiming bitwise replicability unrealistic for a short-term research project, but it is not an insurmountable barrier on a longer time scale. All we need are tighter specifications and implementations that respect them. That's a lot of work, but not a technical challenge. We know how to do it, but we are not (yet) willing to invest the effort to make it happen.
The main reason why I consider bitwise replicability important is software testing. No matter what precise approach is used for testing, it always involves comparing results of computations, either to a known good result, or to the result of another, presumably more reliable, computation. For any application of computing other than number crunching, comparing results means testing for equality, at the bit level. The results are equal or they aren't. If they aren't, there's a reason. You have to figure out what that reason is, and fix the problem.
If you accept the idea that floating-point operations are only approximate, the notion of a computation having one and only one result disappears, and testing becomes impossible. If two computations lead to similar but slightly different results, how do you decide if this is due to a bug or to some "inevitable" fuzziness of floating-point arithmetic? The answer is that you can't. If you accept that bitwise replicability is not possible, you also accept that rigorous software testing is not possible. For some illustrations of this problem, and some interesting discussion around them, see this post on the Software Carpentry blog.
The most common counterargument is that numerical methods are only approximate, that floating-point arithmetic is approximate as well, and that the main source of error comes from these two sources. That may or may not be true in any specific situation, as it really depends on what you are computing. But my point is that this statement can only be true if you assume that the implementation of your method contains no mistakes. The amount of error introduced by a bug in the code is completely unbounded. And even if it's small for some particular test run, it can be very large elsewhere. There is not much point in worrying about the error in an approximate numerical method unless you have some confidence in your code actually implementing this method correctly.
In fact, the common counterargument discussed above conflates several sources of error, which can and should be discussed and analyzed separately. A typical numerical computation is the result of several steps, starting from a mathematical model that takes the form of algebraic or differential equations:
Construct a computable approximation1 to the original equations, using techniques such as discretization of continuous quantities.
Replace real-numbers by floating-point numbers.
Implement the floating-point version in software.
The errors introduced in the first step are the subject of numerical analysis, a well-established domain of applied mathematics. They are well understood for most commonly employed numerical methods. The errors introduced in the second step are rarely discussed explicitly, outside of a small circle of researchers interested in the peculiarities of floating-point arithmetic. The third step should not introduce any errors, and that should be verified by testing. But uncoupling steps 2 and 3 is possible only if our software tools guarantee bitwise replicability.
So why don't today's tools permit this? The reason is a mixture of widespread ignorance about floating-point arithmetic and the desire to get maximum performance. Both come into play in step 2, which is approximating discrete equations for real numbers by discrete equations for floating-point numbers. Most scientific programmers are unaware that this is an approximation that they should understand and control. They just type their real-number equation into a program and expect the computer to handle it somehow. Compiler writers and language specification authors take advantage of this ignorance and declare this step their business, profiting from the many optimization possibilities it offers.
The optimization opportunities come from the fact that a typical real-number equation has a large number of a priori equally plausible floating-point number approximations. Many of the identities for real numbers do not apply to floating-point numbers, for example associativity of addition and multiplication. Where the real-number equation says a+b+c, there are three floating-point approximations: (a+b)+c, a+(b+c), and (a+c)+b. For more complex equations, the number of variants quickly becomes important. The results of these variants are not the same, but which one to choose? The choice should be made after a careful analysis of the relative precision and performance of each variant. There should be tool support to help with this. But what happens in practice, most of the time, is that the choice is made by the compiler, which goes exclusively for performance. Since every compiler optimizes differently, the same program source code yields different results on different platforms. And that's why we don't have bitwise replicability.
To prevent any misunderstanding: I am not saying that production-level compiled code needs to ensure bitwise reproducibility across machines. It's OK to have compiler optimization options that introduce platform-specific approximations. But it should be possible to reproduce one unique result identically on all platforms. This result is then the reference against which additional "lossy" optimizations can be tested.
Footnotes:
1 I am using the term "computable approximation" somewhat vaguely here. While the original continuous-variable equations are almost always non-computable, and the numerical approximations are mostly computable, there are exceptions on both sides. The main focus of numerical analysis is not computability in the strict sense of computability theory, but "practical" computability that has the subsequent transformation to floating-point operations in mind.
The importance of reproducibility in computational science is being more and more recognized, which I think is a good sign. However, I also notice a lot of confusion about what reproducibility means exactly, and also confusion about the difference (if any) between reproducibility and replicability. I don't see a consensus yet about the exact meaning of these terms, but I would like to give my own definitions and justify them by putting them into the general context of computational science.
I'll start with the concept of reproducibility as it was used in science long before computers even existed. It refers to the reproducibility of the conclusions of a scientific study. These conclusions can take very different forms depending on the question that was being explored. It can be a simple "yes" or "no", e.g. in answering questions such as "Is the gravitational force acting in this stone the same everywhere on the Earth's surface?" or "Does ligand A bind more strongly to protein X than ligand B?" It can also be a number, as in "What is the lattice energy of NaCl?", or a mathematical function, as in "How does a spring's restoring force vary with elongation?" Any such result should come with an estimation of its precision, such as an error bar on numbers, or a reliability estimate for a yes/no answer. Reproducing a scientific conclusion means finding a "close enough" answer by performing "similar" experiments and analyses. As the terms "close enough" and "similar" show, reproducibility involves human judgement, which may well evolve over time. Reproducibility is thus not an absolute feature of a specific result, but the evaluation of a result in the context of the current state of knowledge and technology in a scientific domain. Every attempt to reproduce a given result independently (different people, tools, methods, …) augments scientific knowledge: If the reproduction leads to a "close enough" results, it provides information about the precision with which the results can be obtained, and if if doesn't, it points to some previously unrecognized crucial difference between the two experiments, which can then be explored.
Replication refers to something much more specific: repeating the exact steps in an experiment using the same (or equivalent) equipment, and comparing the outcomes. Replication is part of testing an experimental setup, or a form of quality assurance. If I measure the same quantity ten times using the same equipment and experimental samples, and get ten slightly different values, then I can use these numbers to estimate the precision of my equipment. If that precision is not sufficient for the purposes of my planned scientific study, then the equipment is not suitable.
It is useful to describe the process of doing research by a two-layer model. The fundamental layer is the technology layer: equipment and procedures that are well understood and whose precision is known from many replication attempts. On top of this, there is the research layer: the well-understood equipment is used in order to obtain new scientific information and draw conclusions from them. Any scientific project aims at improving one or the other layer, but not both at the same time. When you want to get new scientific knowledge, you use trusted equipment and procedures. When you want to improve the equipment or the procedures, you do so by doing test measurements on well-known systems. Reproducibility is a concept of the research layer, replicability belongs to the technology layer.
All this carries over identically to computational science, in principle. There is the technology layer, consisting of computers and the software that runs on them, and the research layer, which uses this technology to explore theoretical models or to interpret experimental data. Replicability belongs to the technology level. It increases trust in a computation and thus its components (hardware, software, overall workflow, provenance tracking, …). If a computation cannot be replicated, then this points to some kind of problem:
different input data that was not recorded in the workflow (interactive user input, a random number stream initialized from the current time, …)
a bug in the software (uninitialized variables, compiler bugs, …)
a fault in the hardware (an unreliable memory chip, a design flaw in the processor, …)
an ambiguous specification of the result of the computation
Ideally, the non-replicability should be eliminated, but at the very least its cause should be understood. This turns out to be very difficult in practice, in today's computing environments, essentially because case 4 is frequent and hard to avoid (today's popular programming languages are ambiguous), and because case 4 makes it impossible to identify cases 2 and 3 with certainty. I see this as a symptom of the immaturity of today's computing environments, which the computational science community should aim to improve on. The technology for removing case 4 exists. The keyword is "formal methods", and there are first attempts to apply them to scientific computing, but this remains an exotic approach for now.
As in experimental science, reproducibility belongs to the research layer and cannot be guaranteed or verified by any technology. In fact, the "reproducible research" movement is really about replicability - which is perhaps one reason for the above-mentioned confusion.
There is at the moment significant disagreement about the importance of replicability. At one end of the spectrum, there is for example Ian Gent's recomputation manifesto, which stresses the importance of replicability (which in the context of computational science he calls recomputability) because building on past work is possible only if it can be replicated as a first step. At the other end, Chris Drummond argues that replicability is "not worth having" because it doesn't contribute much to the real goal, which is reprodcucibility. It is worth reading both of these papers, because they both do a very good job at explaining their arguments. There is actually no contradiction between the two lines of arguments, the different conclusions are due to different criteria being applied: Chris Drummond sees replicability as valuable only if it improves reproducibility (which indeed it doesn't), whereas Ian Gent sees value in it for a completely different reason: it makes future research more efficient. Neither one mentions the main point in favor of replicability that I have made above: that replicability is a form of quality assurance and thus increases trust in published results.
It is probably a coincidence that both of the papers cited above use the term "computational experiment", which I think should best be avoided in this context. In the natural sciences, the term "experiment" traditionally refers to constructing a setup to observe nature, which makes experiments the ultimate source of truth in science. Computations do not have this status at all: they are applications of theoretical models, which are always imperfect. In fact, there is an interesting duality between the two: experiments are imperfect observations of the ultimate truth, whereas computations are, in the absence of buggy or ambiguous software, perfect observations of the consequences of imperfect models. Using the same term for these two concepts is a source of confusion, as I have pointed out earlier.
This fundamental difference between experiments and computations also means that replicability has a different status in experimental and computational science. When doing imperfect observations of nature, evaluating replicability is one aspect of evaluating the imperfection of the observation. Perfect observation is impossible, both due to technological limitations and for fundamental reasons (any observation modifies what is being observed). On the other hand, when computing the consequences of imperfect models, replicability does not measure the imperfections of the model, but the imperfections of the computation, which can theoretically be eliminated.
The main source of imperfections in computations is the complexity of computer software (considering the whole software stack, from the operating system to the scientific software). At this time, it is not clear if we will ever succeed in taming this complexity. Our current digital computers are chaotic systems, in which even the tiniest change (flipping a bit in memory, or replacing a single character in a program source code file) can change the result of a computation beyond any bounds. Chaotic behavior is clearly an undesirable feature in any scientific equipment (I can't think of any experimental apparatus suffering from it), but for computation we currently have no other choice. This makes quality assurance techniques, including replicability but also more standard software engineering practices such as unit testing, all the more important if we want computational results to be trustworthy.
The other day I was looking at the release notes for the recently published release 1.8 of NumPy, the library that is the basis for most of the Scientific Python ecosystem. As usual, it contains a list of new features and improvements, but also sections such as "dropped support" (for Python 2.4 and 2.5) and "future changes", to be understood as "incompatible changes that you should start to prepare for". Dropping support for old Python releases is understandable: maintaining compatibility and testing it is work that needs to be done by someone, and manpower is notoriously scarce for projects such as NumPy. Many of the announced changes are in the same category: they permit removing old code and thus reduce maintenance effort. Other announced changes have the goal of improving the API, and I suppose they were more controversial than the others, as it is rarely obvious that one API is better than another one.
From the point of view of reproducible research, all these changes are bad news. They mean that libraries and scripts that work today will fail to work with future NumPy releases, in ways that their users, who are usually not the authors, cannot easily understand or fix. Actively maintained libraries will of course be adapted to changes in NumPy, but much, perhaps most, scientific software is not actively maintained. A PhD student doing computational reasearch might well publish his/her software along with the thesis, but then switch subjects, or leave research altogether, and never look at the old code again. There are also specialized libraries developed by small teams who don't have the resources to do as much maintenance as they would like.
Of course NumPy is not the only source of instability in the Python platform. The most visible change in the Python ecosystem is the evolution of Python itself, whose 3.x series is not compatible with the initial Python language. It is difficult to say at this time for how long Python 2.x will be maintained, but it is well possible that much of today's scientific software written in Python will become difficult to run ten years from now.
The problem of scientific publications becoming more and more difficult to use is not specific to computational science. A theoretical physicist trying to read Isaac Newton's works would have a hard time, because the mathematical language of physics has changed considerably over time. Similarly, an experimentalist trying to reproduce Galileo Galilei's experiments would find it hard to follow his descriptions. Neither is a problem in practice, because the insights obtained by Newton and Galilei have been reformulated many times since then and are available in today's language in the form of textbooks. Reading the original works is required only for studying the history of science. However, it typically takes a few decades before specific results are universally recognized as important and enter the perpetually maintained canon of science.
The crucial difference with computations is that computing platforms evolve much faster than scientific research. Researchers in fields such as physics and chemistry routinely consult original research works that are up to thirty years old. But scientific software from thirty years ago is almost certainly unusable today without changes. The state of today's software thirty years from now is likely to be worse, since software complexity has increased significantly. Thirty years ago, the only dependencies a scientific program would have is a compiler and perhaps one of a few widely known numerical libraries. Today, even a simple ten-line Python script has lots of dependencies, most of the indirectly through the Python interpreter.
One popular attitude is to say: Just run old Python packages with old versions of Python, NumPy, etc. This is an option as long as the versions you need are recent enough that they can still be built and installed on a modern computer system. And even then, the practical difficulties of working with parallel installation of multiple versions of several packages are considerable, in spite of tools designed to help with this task (have a look at EasyBuild, hashdist, conda, and Nix or its offshoot Guix).
An additional difficulty is that the installation instructions for a library or script at best mention a minimum version number for dependencies, but not the last version with which they were tested. There is a tacit assumption in the computing world that later versions of a package are compatible with earlier ones, although this is not true in practice, as the example of NumPy shows. The Python platform would be a nicer place if any backwards-incompatible change were accompanied by a change in package name. Dependencies would then be evident, and the different incompatible versions could easily be installed in parallel. Unfortunately this approach is rarely taken, a laudable exception being Pyro, whose latest incarnation is called Pyro4 to distinguish it from its not fully compatible predecessors.
I have been thinking a lot about this issue recently, because it directly impacts my ActivePapers project. ActivePapers solves the dependency versioning problem for all code that lives within the ActivePaper universe, by abandoning the notion of a single collection of "installed packages" and replacing it by explicit references to a specific published version. However, the problem persists for packages that cannot be moved inside the ActivePaper universe, typically because of extension modules written in a compiled language. The most fundamental dependencies of this kind are NumPy and h5py, which are guaranteed to be available in an ActivePapers installation. ActivePapers does record the version numbers of NumPy and h5py (and also HDF5) that were used for each individual computation, but it has currently no way to reproduce that exact environment at a later time. If anyone has a good idea for solving this problem, in a way that the average scientist can handle without becoming a professional systems administrator, please leave a comment!
As I have pointed out in an earlier post, long-term reproducibility in computational science will become possible only if the community adopts a stable code representation, which needs to be situated somewhere in between processor instruction sets and programming languages, since both ends of this spectrum are moving targets. In the meantime, we will have to live with workarounds.
Today I have published the first release of ActivePapers for Python, available on PyPI or directly from the Mercurial repository on Bitbucket. The release coincides with the publication of my first scientific paper for which the complete code and data is in the supplementary material, available through the J. Chem. Phys. Web site or from Figshare. There is a good chance that this is the first fully reproducible paper in the field of biomolecular simulation, but it is of course difficult to verify such a claim.
ActivePapers is a framework for doing and publishing reproducible research. An ActivePaper is a file that contains code (Python modules and scripts) and data (HDF5 datasets), plus the dependency information between all these pieces. You can change a script and re-run all the computations that depend on it, for example. Once your project is finished, you can publish the ActivePaper as supplementary material to your standard paper. You can also re-use code and data from a published ActivePaper by using DOI-based links, although for the moment this works only for ActivePapers stored on Figshare.
I consider this first release of ActivePapers quite usable (I use it, after all), but it's definitely for "early adopters". You should be comfortable working with command-line tools, for example, and of course you need some experience with writing Python scripts if you want to create your own ActivePaper. For inspecting data, you can use any HDF5-based tool, such as HDFView, though this makes sense only for data that generic tools can handle. My first published ActivePaper contains lots of protein structures, which HDFView doesn't understand at all. I expect tool support for ActivePapers to improve significantly in the near future.
This post was motivated by Ian Gent's recomputation manifesto and his blog post about it. While I agree with pretty much everything said there, there is one point that I strongly disagree with, and here I'd like to explain the reasons in some detail. The point in question is "The only way to ensure recomputability is to provide virtual machines". To be fair, the manifesto specifies that it's the only way "at least for now", so perhaps our disagreement is not as pronounced as it may seem.
I'll start with a quote from the manifesto that shows that we have similar ideas of the time scales over which computational research should be reproducible: "It may be true that code you make available today can be built with only minor pain by many people on current computers. That is unlikely to be true in 5 years, and hardly credible in 20."
So the question is: how can we best ensure that the software used in our computational studies can still be run, with reasonable effort, 20 years from now. To answer that question, we have to look at the possible platforms for computational research.
By a "platform", I mean the combination of hardware and software that is required to use a given piece of digital information. For example, Flash video requires a Flash player and a computer plus operating system that the Flash player can run on. That's what defines the "Flash platform". Likewise, today's "Web platform" (a description that requires a date stamp to be precise, because Web standards evolve so quickly) consists of HTML5, JavaScript, and a couple of related standards. If you want to watch a Flash video in 20 years, you will need a working Flash platform, and if you want to use an archived copy of a 2013 Web site, you need the 2013 Web platform.
If you plan to distribute some piece of digital information with the hope that it will make sense 20 years from now, you must either have confidence in the longevity of the platform, or be willing and able to ensure its long-term maintenance yourself. For the Flash platform, that means confidence in Adobe and its willingness to keep Flash alive (I wouldn't bet on that). For the 2013 Web platform, you may hope that its sheer popularity will motivate someone to keep it alive, but I wouldn't bet on it either. The Web platform is too complex and too ill-defined to be kept alive reliably when no one uses it in daily life any more.
Back to computational science. 20 years ago, most scientific software was written in Fortran 77, often with extensions specific to a machine or compiler. Much software from that era relied on libraries as well, but they were usually written in the same language, so as long as their source code remains available, the platform for all that is a Fortran compiler compatible with the one from back then. For standard Fortran 77, that's not much of a problem, whereas most of the vendor-specific extensions have disappeared since. Much of that 20-year-old software can in fact still be used today. However, reproducing a computational study based on that software is a very different problem: it also requires all the input data and an executable description of the computational protocol. Even in the rare case that all that information is available, it is likely to depend on lots of other software pieces that may not be easy to get hold of any more. The total computational platform for a given research project is in fact as ill-defined as the 2013 Web platform.
Today's situation is worse, because we use more diverse software written in more different languages, and also use more interactive software whose use is notoriously non-reproducible. The only aspect where we have gained in standardization is the underlying hardware and OS layer: pretty much all computational science is done today on x86 processors running Linux. Hence the idea of conserving the full operating environment in the form of a virtual machine. Just fire up VirtualBox (or one of the other virtual machine managers) and run an exact copy of the original study's work environment.
But what is the platform required to run today's virtual machines? It's VirtualBox, or one of its peers. Note however that it's not "any of today's virtual machine managers" because compatibility between their virtual machine formats is not perfect. It may work, or it may not. For simplicity I will use VirtualBox in the following, but you can substitute another name and the basic arguments still hold.
VirtualBox is a highly non-trivial piece of software, and it has very stringent hardware requirements. Those hardware requirements are met by the vast majority of today's computing equipment used in computational science, but the x86 platform is losing market share rapidly on the wider computing device market. VirtualBox doesn't run on an iPad, for example, and probably it never will. Is VirtualBox likely to be around in 20 years? I won't dare a prediction. If x86 survives for another 20 years AND if Oracle sees a continuing interest in this product, then it will. I won't bet on it though.
What we really need for long-term recomputability is a simple platform. A platform that is simple enough that the scientific community alone can afford to keep it alive for its own needs, even if no one else in the world cares about it.
Unfortunately there is no suitable platform today, to the best of my knowledge. Which is why virtual machines are perhaps the best option right now, for lack of a satisfactory one. But if we care about recomputability, we should design and develop a good supporting platform, starting as soon as possible.
For a more detailed discussion of this issue, see this paper written by yours truly. It comes to the conclusion that the closest existing approximation to a good platform is the Java virtual machine. What we'd want ideally is something similar to the JVM, but designed and optimized for scientific applications. A basic JVM implementation is quite simple (the complex JIT stuff is not a requirement), a few orders of magnitude simpler than VirtualBox, and it has no specific hardware dependencies. It's even simpler than many of today's scientific software packages, so the scientific community can definitely afford to keep it alive, The tough part is... no, it's not designing or writing the required software, it's agreeing on a specification. Perhaps it will never happen. Perhaps virtual machines will remain the best choice for lack of a satisfactory one. Or perhaps we will end up compiling our software to asm.js and run in the browser, just because someone else will keep that platform alive for us, no matter how ill-adapted it is to our needs. But don't say you haven't been warned.
In an earlier post, I mentioned the Nix package management system as a candidate for ensuring reproducibility in computational science. What distinguishes Nix from the better known package managers (Debian, RPM, ...) is that it permits the installation of different versions of the same package in parallel, with a dependency tracking system that refers to a precise version of everything, including the versions of the development tools (compilers, ...) that were used to build the libraries and executables. Nix thus remembers for each package the complete details of how it can be reconstructed, which is what we would like to see for ensuring reproducibility.
There are, however, two caveats. First of all, Nix was designed for software installation management and not for computation. While in principle one could define the results (figures, tables, datasets) of some computation as a Nix package and perform the computation by installing the package, such an approach is quite cumbersome with the Nix support tools designed with a different task in mind. However, computation-specific support tools would probably suffice to fix this. Second, while the design of Nix looks quite sound, it is a young project with much less manpower behind it than the big package managers of the Linux world. This means there are fewer package definitions and they are overall less reliable. For example, I haven't yet managed to install my research computing environment (Python, NumPy, matplotlib, plus a few more packages) using Nix under MacOS X, because some packages simply fail to build. Again this is not an insurmountable problem, but it requires some serious effort to fix.
The Nix documentation is pretty good at describing how to use the package manager and the collection of package definitions for Linux and MacOS X named Nixpkgs. It is not so good at giving a basic understanding of how Nix works, which becomes important when you want to use it for something else than traditional package management. The following overview is the result of my own explorations of Nix. I am not a Nix authority, so be warned that there may be mistakes or misunderstandings.
At the heart of Nix is the "Nix store", a central database where everything managed by Nix is kept. Its default location is /nix/store and if you look at it you see an overwhelmingly long list of crypic filenames. Let's zoom in on something to see what's going on. Here is what ls -l /nix/store/*zlib* shows on my machine:
-r--r--r-- 1 hinsen staff 1000 Jan 1 1970 /nix/store/12vkkhs36xffzpqjaaa3vqhqv2yc97vs-zlib-1.2.6.drv -r--r--r-- 1 hinsen staff 1181 Jan 1 1970 /nix/store/gymcn145ihhmymm6yk2wxqfd49s5dzdq-zlib-1.2.6.drv dr-xr-xr-x 5 hinsen staff 170 Jan 1 1970 /nix/store/mrdqnzzr80rkfnm59q6aywdba6776f66-zlib-1.2.6 -r--r--r-- 1 hinsen staff 1000 Jan 1 1970 /nix/store/sj8l48kfc40wh8adb5pa843lwy38hskb-zlib-1.2.6.drv -r--r--r-- 1 hinsen staff 1686 Jan 1 1970 /nix/store/xpm2xja2zv5agmdzgi362jqd5xx9ny10-zlib-1.2.6.tar.gz.drv
The single directory in that list actually contains the zlib installation in the familiar Unix file layout that you find under /usr or /usr/local:
~> ls -R /nix/store/mrdqnzzr80rkfnm59q6aywdba6776f66-zlib-1.2.6 /nix/store/mrdqnzzr80rkfnm59q6aywdba6776f66-zlib-1.2.6: include lib share
Note that it contains just zlib, and nothing else, in particular not zlib's dependencies. Each library or application has its own directory in the Nix store.
Next, let's look at all the other files, those with the extension .drv (for "derivation", a Nix term for any artefact derived from human-provided input). There are three files that end in zlib-1.2.6.drv and one that ends in zlib-1.2.6.tar.gz.drv. Let's look at the contents of the last one first. I have made it more readable by adding whitespace:
If that looks like a computational expression in a programming language, that's because it is. Don't worry, it's not something you are expected to write yourself, these expressions are created from the package definitions written in a more user-friendly syntax called "Nix expressions", which is very well documneted in the Nix documentation.. The expression shown above defines how to make (or "realise" in Nix jargon) the derivation /nix/store/s9qgdh7g22nx433y3lk62igm5zh48dxj-zlib-1.2.6.tar.gz, which is a rather simple one because the file is simply downloaded and verified for a known checksum. But even such a simple derivation has dependencies: the "standard environment" stdenv and the list of download mirror sites, mirrors-list.
It's time to say something about those funny 32-character prefixes in all the file names in the Nix store. You may have noticed that the zlib file list above contains two entries for zlib-1.2.6.drv that are identical except for this prefix. It looks as if the prefix is there to distinguish things that would otherwise be identical. This is true, and the information encoded in the prefix (which is a hash code) is the complete set of dependencies. The two zlib derivations differ in the version of the standard environment they were built with. I have both of these in my Nix store because I have played around with different releases of Nixpkgs. Nix really tries to keep track of every single dependency, including the exact versions of the various tools (mainly compilers) that were used in building a binary installation. That means you can keep lots of different versions of every single item on your system at the same time, and trace back exactly how they were built. You can also send a copy of the relevant derivation files (those with the .drv extension) to someone else, who can reproduce the exact same environment by "realising" those derivations again.
With so many zlibs floating around, which one does Nix use when you ask it to install some application that uses zlib? The one you specify. When some application requires zlib as a dependency, you have to tell Nix exactly which zlib derivation you want to be used. You don't normally do this manually for every single build (though you could), you'd rather use a coherent set of package definitions (such as Nixpkgs) that specifies all the interdependencies among hundreds of packages. The package definitions take the form of "Nix expressions", which are written in a language specifically designed for this purpose. Files containing Nix expressions have the extension .nix. Since the language is rather well documented in the Nix manual, I won't say any more about it here. A good starting point is to explore Nixpkgs. It helps to know that the central file is pkgs/top-level/all-packages.nix. This file imports the definitions of individual packages from their respective packages and makes a consistent package collection from them. When you build a particular derivation from Nixpkgs, only the packages listed explicitly as its dependencies are available in the build environment that is set up specifically for this build operation. No "default library" (such as /usr/lib) is used at all.
There is one more layer to Nix, whose role is twofold: making it convenient for users to work with programs installed through Nix, and pemitting to remove packages that were installed but are no longer needed. Let's start with the second aspect because it is the simpler one: packages can be removed as soon as nobody needs them any more. This requires a way to figure out which packages are still needed. Obviously the packages that some user on the system wants to access are "needed", and that's why cleanup is related to user profiles which I will cover in a minute. The remaining needed packages are the dependencies of other needed packages. So once we know the packages that all users put together request to use, we can figure out which packages can safely be deleted. This clean-up operation is called "garbage collection" and handled by the command nix-store --gc.
Nix user environments are managed using the command nix-env, and if you don't care about how Nix works, that command is the only one you may ever need. Each user has his/her own environment, of course, which consists mainly of a directory named $HOME/.nix-profile. That directory contains subdirectories called bin, lib, man etc. whose names should sound familiar. They contain nothing but symbolic links into the Nix store. These links define which package the user actually accesses, by putting $HOME/.nix-profile/bin on th3 PATH environment variable. When you use nix-env to install a package, Nix builds it and puts it into the Nix store (unless it's already there), and then creates symbolic links in your Nix profile, which may replace links to some different version of a package. It is important to understand that your use profile never enters into the build process of any Nix derivation. Your profile is exclusively for your own use and has no impact on Nix package management other than protecting the packages you use from being removed during garbage collection.
So far for a first report on my exploration of Nix. I will continue trying to get my computational environment built with Nix, so that I can start to explore how to use it for reproducible computations. Watch this space for news.
PS: After I published this post initially, the friendly people on the Nix mailing list pointed out some additional material for learning about Nix. First of all, there is Eelco Dolstra's thesis entitled "The Purely Functional Software Deployment Model", which is what you should read if you really want to know everything about Nix. There's also Sander van der Burg's blog which has some very detailed posts about Nix and what it can be used for. You could start with this introduction.
When the Greek philosopher Heraclitus pronounced his famous "πάντα ῥεῖ" (everything flows), he most probably was not thinking about software. But it applies to software as much as to other aspects of life: software is in perpetual change, being modified to remove bugs, add features, and adapt it to changing environments. The management of change is now a well-established part of software engineering, with the most emblematic tool being version control. If you are developing software without using version control, stop reading this immediately and learn about Mercurial or Git, the two best version control systems available today. That's way more important than reading the rest of this post.
Software developers use version control to keep track of the evolution of their software, to coordinate team development, and to manage experimental features. But version control is also of interest for software users: it permits them to refer to a specific version of a piece of software they use in a unique and reproducible way, even if that version is not the current one, nor perhaps even an official numbered release. In fact, official numbered releases are becoming a relict of the past. They make little sense in an Open Source universe where everyone has access to source code repositories under version control. In that situation, an official release is nothing but a bookmark pointing to a specific commit number. There is no need for a release number.
Why would you want to refer to a specific version of a piece of software, rather than always use the latest one? There are many reasons. As software evolves, some bugs get fixed but others sneak in. You may prefer the bugs you know to the ones that could surprise you. Sometimes later versions of some software are not fully compatible with their predecessors, be it by design or by mistake. And even if you want to use the very latest version at any time, you might still want to note which version you used for a specific application. In scientific computing, this is one of the fundamental principles of reproducible research: note carefully, and publish, the exact versions of all pieces of software that were used for obtaining any published research result. It's the only way for you and others to be able to understand exactly what happened when you look at your work many years later.
Another undeniable reality of modern software, in particular in the Open Source universe, is that it's modular. Developers use other people's software, especially if it's well written and has the reputation of being reliable, rather than reinventing the wheel. The typical installation instructions of a piece of Open Source software start with a list of dependencies, i.e. packages you have to install before you can install the current one. And of course the packages in the dependency list have their own dependency list. The number of packages to install can be overwhelming. The difficulties of dependency management are so widespread that the term "dependency hell" has been coined to refer to them.
Systems programmers have come up with a solution to that problem as well: dependency management tools, better known as package managers. Such tools keep a database of what is installed and which package depends on which other ones. The well-known Linux distributions are based on such package managers, of which the ones developed by Debian and RedHat are the most popular ones and are now used by other distributions as well. For MacOS X, MacPorts and Fink are the two most popular package managers, and I suspect that the Windows world has its own ones.
One of the major headaches that many computer users face is that version management and dependency management don't cooperate. While most package managers permit to state a minimal version number for a dependency, they don't permit to prescribe a precise version number. There is a good reason for this: the way software installation is managed traditionally on Unix systems makes it impossible to install multiple versions of the same package in parallel. If packages A and B both depend on C, but require different versions of it, there is simply no simple solution. Today's package managers sweep this problem under the rug and pretend that higher version numbers are always as least as good as their predecessors. They will therefore install the higher of the two version numbers required by A and B, forcing one of them to use a version different from its preference.
Anyone who has been using computers intensively for a few years has probably run into such a problem, which manifests itself by some program not working correctly any more after another one, seemingly unrelated, has been installed. Another variant is that an installation fails because some dependency is available in a wrong version. Such problems are part of "dependency hell".
This situation is particularly problematic for the computational scientist who cares about the reproducibility of computed results. At worst, verifying results from 2005 by comparing to results from 2009 can require two completely separate operating system installations running in separate virtual machines. Under such conditions, it is difficult to convince one's colleagues to adopt reproducible research practices.
While I can't propose a ready-to-use solution, I can point out some work that shows that there is hope for the future. One interesting tool is the Nix package manager, which works much like the package managers by Debian or RedHat, but permits installing multiple versions of the same package in parallel, and registers dependencies with precise versions. It could be used as a starting point for managing software for reproducible research, the main advantage being that it should work with all existing software. The next step would be to make each result dataset or figure a separate "package" whose complete dependency list (software and datasets) is managed by Nix with references to precise version numbers. I am currently exploring this approach; watch this space for news about my progress.
For a system even better suited to the needs of reproducible computational science, I refer to my own ActivePapers framework, which combines dependency management and version control for code and data with mechanisms for publishing code+data+documentation packages and re-use code from other publications in a secure way. I have to admit that it has a major drawback as well: it requires all code to run on the Java Virtual Machine (in order to guarantee portability and secure execution), which unfortunately means that most of today's scientific programs cannot be used. Time will tell if scientific computing will adopt some virtual machine in the future that will make such a system feasible in real life. Reproducible research might actually become a strong argument in favour of such a development.