diff options
| -rw-r--r-- | paper.tex | 1089 | ||||
| -rw-r--r-- | tex/src/appendix-existing-solutions.tex | 443 | ||||
| -rw-r--r-- | tex/src/appendix-existing-tools.tex | 537 | ||||
| -rw-r--r-- | tex/src/appendix-necessity.tex | 79 |
4 files changed, 1061 insertions, 1087 deletions
@@ -683,1093 +683,8 @@ The Pozna\'n Supercomputing and Networking Center (PSNC) computational grant 314 \else \clearpage \appendices -\section{Necessity for reproducible research} -\label{sec:introduction} - -The increasing volume and complexity of data analysis has been highly productive, giving rise to a new branch of ``Big Data'' in many fields of the sciences and industry. -However, given its inherent complexity, the mere results are barely useful alone. -Questions such as these commonly follow any such result: -What inputs were used? -What operations were done on those inputs? How were the configurations or training data chosen? -How did the quantitative results get visualized into the final demonstration plots, figures or narrative/qualitative interpretation? -Could there be a bias in the visualization? -See Figure \ref{fig:questions} for a more detailed visual representation of such questions for various stages of the workflow. - -In data science and database management, this type of metadata is commonly known as \emph{data provenance} or \emph{data lineage}. -Data lineage is being increasingly demanded for integrity checking from both the scientific, industrial and legal domains. -Notable examples in each domain are respectively the ``Reproducibility crisis'' in the sciences that was reported by \emph{Nature} after a large survey \citeappendix{baker16}, and the General Data Protection Regulation (GDPR) by the European Parliament and the California Consumer Privacy Act (CCPA), implemented in 2018 and 2020, respectively. -The former argues that reproducibility (as a test on sufficiently conveying the data lineage) is necessary for other scientists to study, check and build-upon each other's work. -The latter requires the data intensive industry to give individual users control over their data, effectively requiring thorough management and knowledge of the data lineage. -Besides regulation and integrity checks, having robust data governance (management of data lineage) in a project can be very productive: it enables easy debugging, experimentation on alternative methods, or optimization of the workflow. - -In the sciences, the results of a project's analysis are published as scientific papers, which have traditionally been the primary conveyor of the lineage of the results: usually in narrative form, especially within the ``Methods'' section of the paper. -From our own experience, this section is often that which is the most intensively discussed during peer review and conference presentations, showing its importance. -After all, a result is defined as ``scientific'' based on its \emph{method} (the ``scientific method''), or lineage in data-science terminology. -In industry, however, data governance is usually kept as a trade secret and is not published openly or widely scrutinized. -Therefore, the main practical focus here will be in the scientific front, which has traditionally been more open to the publication of methods and anonymous peer scrutiny. - -\begin{figure*}[t] - \begin{center} - \includetikz{figure-project-outline}{width=\linewidth} - \end{center} - \caption{\label{fig:questions}Graph of a generic project's workflow (connected through arrows), highlighting the various issues and questions on each step. - The green boxes with sharp edges are inputs and the blue boxes with rounded corners are the intermediate or final outputs. - The red boxes with dashed edges highlight the main questions at various stages in the workchain. - The orange box, surrounding the software download and build phases, lists some commonly recognized solutions to the questions in it; for morediscussion, see Appendix \ref{appendix:independentenvironment}. - } -\end{figure*} - -The traditional format of a scientific paper has been very successful in conveying the method and the results during recent centuries. -However, the complexity mentioned above has made it impossible to describe all the analytical steps of most modern projects to a sufficient level of detail. -Citing this difficulty, many authors limit themselves to describing the very high-level generalities of their analysis, while even the most basic calculations (such as the mean of a distribution) can depend on the software implementation. - -Due to the complexity of modern scientific analysis, a small deviation in some of the different steps involved can lead to significant differences in the final result. -Publishing the precise codes of the analysis is the only guarantee of allowing this to be investigated. -For example, \citeappendix{smart18} describes how a 7-year old conflict in theoretical condensed matter physics was only identified after the different groups' codes were shared. -Nature is already a black box that we are trying hard to unlock and understand. -Not being able to experiment on the methods of other researchers is an artificial and self-imposed black box, wrapped over the original, and wasting much of researchers' time and energy. - -A dramatic example showing the importance of sharing code is \citeappendix{miller06}, in which a mistaken flipping of a column was discovered, leading to the retraction of five papers in major journals, including \emph{Science}. -Ref.\/ \citeappendix{baggerly09} highlighted the inadequate narrative description of analysis in several papers and showed the prevalence of simple errors in published results, ultimately calling their work ``forensic bioinformatics''. -References \citeappendix{herndon14} and \citeappendix{horvath15} also reported similar situations and \citeappendix{ziemann16} concluded that one-fifth of papers with supplementary Microsoft Excel gene lists contain erroneous gene name conversions. -Such integrity checks are a critical component of the scientific method, but are only possible with access to the data and codes and \emph{cannot be resolved from analysing the published paper alone}. - -The completeness of a paper's published metadata (or ``Methods'' section) can be measured by a simple question: given the same input datasets (supposedly on a third-party database like \href{http://zenodo.org}{zenodo.org}), can another researcher reproduce the exact same result automatically, without needing to contact the authors? -Several studies have attempted to answer this with different levels of detail. -For example \citeappendix{allen18} found that roughly half of the papers in astrophysics do not even mention the names of any analysis software they have used, while \cite{menke20} found that the fraction of papers explicitly mentioning their software tools has greatly improved in medical journals over the last two decades. - -Ref.\/ \citeappendix{ioannidis2009} attempted to reproduce 18 published results by two independent groups, but only fully succeeded in two of them and partially in six. -Ref.\/ \citeappendix{chang15} attempted to reproduce 67 papers in well-regarded economic journals with data and code: only 22 could be reproduced without contacting authors, and more than half could not be replicated at all. -Ref.\/ \citeappendix{stodden18} attempted to replicate the results of 204 scientific papers published in the journal Science \emph{after} that journal adopted a policy of publishing the data and code associated with the papers. -Even though the authors were contacted, the success rate was $26\%$. -Generally, this problem is unambiguously felt in the community: \citeappendix{baker16} surveyed 1574 researchers and found that only $3\%$ did not see a ``reproducibility crisis''. - -This is not a new problem in the sciences: in 2011, Elsevier conducted an ``Executable Paper Grand Challenge'' \citeappendix{gabriel11}. -The proposed solutions were published in a special edition. -Some of them are reviewed in Appendix \ref{appendix:existingsolutions}, but most have not been continued since then. -In 2005, Ref.\/ \citeappendix{ioannidis05} argued that ``most claimed research findings are false''. -Even earlier, in the 1990s, Refs \cite{schwab2000}, \citeappendix{buckheit1995} and \cite{claerbout1992} described this same problem very eloquently and provided some of the solutions that they adopted. -While the situation has improved since the early 1990s, the problems mentioned in these papers will resonate strongly with the frustrations of today's scientists. -Even earlier yet, through his famous quartet, Anscombe \citeappendix{anscombe73} qualitatively showed how the distancing of researchers from the intricacies of algorithms and methods can lead to misinterpretation of the results. -One of the earliest such efforts we found was \citeappendix{roberts69}, who discussed conventions in FORTRAN programming and documentation to help in publishing research codes. - -From a practical point of view, for those who publish the data lineage, a major problem is the fast evolving and diverse software technologies and methodologies that are used by different teams in different epochs. -Ref.\/ \citeappendix{zhao12} describes it as ``workflow decay'' and recommends preserving these auxiliary resources. -But in the case of software, this is not as straightforward as for data: if preserved in binary form, software can only be run on a certain operating system on particular hardware, and if kept as source code, its build dependencies and build configuration must also be preserved. -Ref.\/ \citeappendix{gronenschild12} specifically studies the effect of software version and environment and encourages researchers to not update their software environment. -However, this is not a practical solution because software updates are necessary, at least to fix bugs in the same research software. -Generally, software is not a secular component of projects, where one software package can easily be swapped with another. -Projects are built around specific software technologies, and research in software methods and implementations is itself a vibrant research topic in many domains \citeappendix{dicosmo19}. - - - - - - - - - - -\section{Survey of existing tools for various phases} -\label{appendix:existingtools} - -Data analysis workflows (including those that aim for reproducibility) are commonly high-level frameworks which employ various lower-level components. -To help in reviewing existing reproducible workflow solutions in light of the proposed criteria in Appendix \ref{appendix:existingsolutions}, we first need to survey the most commonly employed lower-level tools. - -\subsection{Independent environment} -\label{appendix:independentenvironment} - -The lowest-level challenge of any reproducible solution is to avoid the differences between various run-time environments, to a desirable/certain level. -For example different hardware, operating systems, versions of existing dependencies, and etc. -Therefore any reasonable attempt at providing a reproducible workflow starts with isolating its running envionment from the host environment. -There are three general technologies that are used for this purpose and reviewed below: -1) Virtual machines, -2) Containers, -3) Independent build in host's file system. - -\subsubsection{Virtual machines} -\label{appendix:virtualmachines} -Virtual machines (VMs) host a binary copy of a full operating system that can be run on other operating systems. -This includes the lowest-level operating system component or the kernel. -VMs thus provide the ultimate control one can have over the run-time environment of an analysis. -However, the VM's kernel does not talk directly to the running hardware that is doing the analysis, it talks to a simulated hardware layer that is provided by the host's kernel. -Therefore, a process that is run inside a virtual machine can be much slower than one that is run on a native kernel. -An advantages of VMs is that they are a single file which can be copied from one computer to another, keeping the full environment within them if the format is recognized. -VMs are used by cloud service providers, enabling fully independent operating systems on their large servers (where the customer can have root access). - -VMs were used in solutions like SHARE \citeappendix{vangorp11} (which was awarded second prize in the Elsevier Executable Paper Grand Challenge of 2011 \citeappendix{gabriel11}), or in suggested reproducible papers like \citeappendix{dolfi14}. -However, due to their very large size, these are expensive to maintain, thus leading SHARE to discontinue its services in 2019. -The URL to the VM file \texttt{provenance\_machine.ova} that is mentioned in \citeappendix{dolfi14} is not currently accessible (we suspect that this is due to size and archival costs). - -\subsubsection{Containers} -\label{appendix:containers} -Containers also host a binary copy of a running environment, but do not have their own kernel. -Through a thin layer of low-level system libraries, programs running within a container talk directly with the host operating system kernel. -Otherwise, containers have their own independent software for everything else. -Therefore, they have much less overhead in hardware/CPU access. -Like VMs, users often choose an operating system for the container's independent operating system (most commonly GNU/Linux distributions which are free software). - -Below we review some of the most common container solutions: Docker and Singularity. - -\begin{itemize} -\item {\bf\small Docker containers:} Docker is one of the most popular tools today for keeping an independent analysis environment. - It is primarily driven by the need of software developers for reproducing a previous environment, where they have root access mostly on the ``cloud'' (which is just a remote VM). - A Docker container is composed of independent Docker ``images'' that are built with a \inlinecode{Dockerfile}. - It is possible to precisely version/tag the images that are imported (to avoid downloading the latest/different version in a future build). - To have a reproducible Docker image, it must be ensured that all the imported Docker images check their dependency tags down to the initial image which contains the C library. - - An important drawback of Docker for high performance scientific needs is that it runs as a daemon (a program that is always running in the background) with root permissions. - This is a major security flaw that discourages many high performance computing (HPC) facilities from providing it. - -\item {\bf\small Singularity:} Singularity \citeappendix{kurtzer17} is a single-image container (unlike Docker which is composed of modular/independent images). - Although it needs root permissions to be installed on the system (once), it does not require root permissions every time it is run. - Its main program is also not a daemon, but a normal program that can be stopped. - These features make it much safer for HPC administrators to install compared to Docker. - However, the fact that it requires root access for the initial install is still a hindrance for a typical project: if Singularity is not already present on the HPC, the user's science project cannot be run by a non-root user. - -\item {\bf\small Podman:} Podman uses the Linux kernel containerization features to enable containers without a daemon, and without root permissions. - It has a command-line interface very similar to Docker, but only works on GNU/Linux operating systems. -\end{itemize} - -Generally, VMs or containers are good solutions to reproducibly run/repeating an analysis in the short term (a couple of years). -However, their focus is to store the already-built (binary, non-human readable) software environment. -Because of this they will be large (many Gigabytes) and expensive to archive, download or access. -Recall the two examples above for VMs in Section \ref{appendix:virtualmachines}. But this is also valid for Docker images, as is clear from Dockerhub's recent decision to delete images of free accounts that have not been used for more than 6 months. -Meng \& Thain \citeappendix{meng17} also give similar reasons on why Docker images were not suitable in their trials. - -On a more fundamental level, VMs or contains do not store \emph{how} the core environment was built. -This information is usually in a third-party repository, and not necessarily inside container or VM file, making it hard (if not impossible) to track for future users. -This is a major problem when considering reproducibility which is also highlighted as a major issue in terms of long term reproducibility in \citeappendix{oliveira18}. - -The example of \cite{mesnard20} was previously mentioned in Section \ref{criteria}. -Another useful example is the \href{https://github.com/benmarwick/1989-excavation-report-Madjedbebe/blob/master/Dockerfile}{\inlinecode{Dockerfile}} of \citeappendix{clarkso15} (published in June 2015) which starts with \inlinecode{FROM rocker/verse:3.3.2}. -When we tried to build it (November 2020), the core downloaded image (\inlinecode{rocker/verse:3.3.2}, with image ``digest'' \inlinecode{sha256:c136fb0dbab...}) was created in October 2018 (long after the publication of that paper). -In principle, it is possible to investigate the difference between this new image and the old one that the authors used, but that would require a lot of effort and may not be possible where the changes are not available in a third public repository or not under version control. -In Docker, it is possible to retrieve the precise Docker image with its digest for example \inlinecode{FROM ubuntu:16.04@sha256:XXXXXXX} (where \inlinecode{XXXXXXX} is the digest, uniquely identifying the core image to be used), but we have not seen this often done in existing examples of ``reproducible'' \inlinecode{Dockerfiles}. - -The ``digest'' is specific to Docker repositories. -A more generic/longterm approach to ensure identical core OS components at a later time is to construct the containers or VMs with fixed/archived versions of the operating system ISO files. -ISO files are pre-built binary files with volumes of hundreds of megabytes and not containing their build instructions). -For example the archives of Debian\footnote{\inlinecode{\url{https://cdimage.debian.org/mirror/cdimage/archive/}}} or Ubuntu\footnote{\inlinecode{\url{http://old-releases.ubuntu.com/releases}}} provide older ISO files. - -The concept of containers (and the independent images that build them) can also be extended beyond just the software environment. -For example \citeappendix{lofstead19} propose a ``data pallet'' concept to containerize access to data and thus allow tracing data back wards to the application that produced them. - -In summary, containers or VMs are just a built product themselves. -If they are built properly (for example building a Maneage'd project inside a Docker container), they can be useful for immediate usage and fast moving of the project from one system to another. -With robust building, the container or VM can also be exactly reproduced later. -However, attempting to archive the actual binary container or VM files as a black box (not knowing the precise versions of the software in them, and \emph{how} they were built) is expensive, and will not be able to answer the most fundamental questions. - -\subsubsection{Independent build in host's file system} -\label{appendix:independentbuild} -The virtual machine and container solutions mentioned above, have their own independent file system. -Another approach to having an isolated analysis environment is to use the same filesystem as the host, but installing the project's software in a non-standrard, project-specific directory that does not interfere with the host. -Because the environment in this approach can be built in any custom location on the host, this solution generally does not require root permissions or extra low-level layers like containers or VMs. -However, ``moving'' the built product of such solutions from one computer to another is not generally as trivial as containers or VMs. -Examples of such third-party package managers (that are detached from the host OS's package manager) include Nix, GNU Guix, Python's Virtualenv package and Conda, among others. -Because it is highly intertwined with the way software are built and installed, third party package managers are described in more detail as part of Section \ref{appendix:packagemanagement}. - -Maneage (the solution proposed in this paper) also follows a similar approach of building and installing its own software environment within the the host's file system but without depending on it beyond the kernel. -However, unlike the third party package maneager mentioned above, based on the Completeness criteria above Maneage's package management is not detached from the specific research/analysis project: the instructions to build the full isolated software environment is maintained with the high-level analysis steps of the project, and the narrative paper/report of the project. - - - - - -\subsection{Package management} -\label{appendix:packagemanagement} - -Package management is the process of automating the build and installation of a software environment. -A package manager thus contains the following information on each software package that can be run automatically: the URL of the software's tarball, the other software that it possibly depends on, and how to configure and build it. -Package managers can be tied to specific operating systems at a very low level (like \inlinecode{apt} in Debian-based OSs). -Alternatively, there are third-party package managers which ca be installed on many OSs. -Both are discussed in more detail below. - -Package managers are the second component in any workflow that relies on containers or VMs for an independent environment, and the starting point in others that use the host's file system (as discussed above in Section \ref{appendix:independentenvironment}). -In this section, some common package managers are reviewed, in particular those that are most used by the reviewed reproducibility solutions of Appendix \ref{appendix:existingsolutions}. -For a more comprehensive list of existing package managers, see \href{https://en.wikipedia.org/wiki/List_of_software_package_management_systems}{Wikipedia}. -Note that we are not including package managers that are specific to one language, for example \inlinecode{pip} (for Python) or \inlinecode{tlmgr} (for \LaTeX). - - - -\subsubsection{Operating system's package manager} -The most commonly used package managers are those of the host operating system, for example \inlinecode{apt} or \inlinecode{yum} respectively on Debian-based, or RedHat-based GNU/Linux operating systems, \inlinecode{pkg} in FreeBSD, among many others in other OSes. - -These package managers are tightly intertwined with the operating system: they also include the building and updating of the core kernel and the C library. -Because they are part of the OS, they also commonly require root permissions. -Also, it is usually only possible to have one version/configuration of a software at any moment and downgrading versions for one project, may conflict with other projects, or even cause problems in the OS. -Hence if two projects need different versions of a software, it is not possible to work on them at the same time in the OS. - -When a container or virtual machine (see Appendix \ref{appendix:independentenvironment}) is used for each project, it is common for projects to use the containerized operating system's package manager. -However, it is important to remember that operating system package managers are not static: software are updated on their servers. -Hence, simply running \inlinecode{apt install gcc}, will install different versions of the GNU Compiler Collection (GCC) based on the version of the OS and when it has been run. -Requesting a special version of that special software does not fully address the problem because the package managers also download and install its dependencies. -Hence a fixed version of the dependencies must also be specified. - -In robust package managers like Debian's \inlinecode{apt} it is possible to fully control (and later reproduce) the build environment of a high-level software. -Debian also archives all packaged high-level software in its Snapshot\footnote{\inlinecode{\url{https://snapshot.debian.org/}}} service since 2005 which can be used to build the higher-level software environment on an older OS \citeappendix{aissi20}. -Hence it is indeed theoretically possible to reproduce the software environment only using archived operating systems and their own package managers, but unfortunately we have not seen it practiced in scientific papers/projects. - -In summary, the host OS package managers are primarily meant for the operating system components or very low-level components. -Hence, many robust reproducible analysis solutions (reviewed in Appendix \ref{appendix:existingsolutions}) do not use the host's package manager, but an independent package manager, like the ones below discussed below. - -\subsubsection{Packaging with Linux containerization} -Once a software is packaged as an AppImage\footnote{\inlinecode{\url{https://appimage.org}}}, Flatpak\footnote{\inlinecode{\url{https://flatpak.org}}} or Snap\footnote{\inlinecode{\url{https://snapcraft.io}}} the software's binary product and all its dependencies (not including the core C library) are packaged into one file. -This makes it very easy to move that single software's built product to newer systems. -However, because the C library is not included, it can fail on older systems. -Moreover, these are designed for the Linux kernel (using its containerization features) and can thus only be run on GNU/Linux operating systems. - -\subsubsection{Nix or GNU Guix} -\label{appendix:nixguix} -Nix \citeappendix{dolstra04} and GNU Guix \citeappendix{courtes15} are independent package managers that can be installed and used on GNU/Linux operating systems, and macOS (only for Nix, prior to macOS Catalina). -Both also have a fully functioning operating system based on their packages: NixOS and ``Guix System''. -GNU Guix is based on the same principles of Nix but implemented differencely, so we focus the review here on Nix. - -The Nix approach to package management is unique in that it allows exact dependency tracking of all the dependencies, and allows for multiple versions of a software, for more details see \citeappendix{dolstra04}. -In summary, a unique hash is created from all the components that go into the building of the package. -That hash is then prefixed to the software's installation directory. -As an example from \citeappendix{dolstra04}: if a certain build of GNU C Library 2.3.2 has a hash of \inlinecode{8d013ea878d0}, then it is installed under \inlinecode{/nix/store/8d013ea878d0-glibc-2.3.2} and all software that are compiled with it (and thus need it to run) will link to this unique address. -This allows for multiple versions of the software to co-exist on the system, while keeping an accurate dependency tree. - -As mentioned in \citeappendix{courtes15}, one major caveat with using these package managers is that they require a daemon with root privileges. -This is necessary ``to use the Linux kernel container facilities that allow it to isolate build processes and maximize build reproducibility''. -This is because the focus in Nix or Guix is to create bit-wise reproducible software binaries and this is necessary in the security or development perspectives. -However, in a non-computer-science analysis (for example natural sciences), the main aim is reproducibile \emph{results} that can also be created with the same software version that may not be bitwise identical (for example when they are installed in other locations, because the installation location is hardcoded in the software binary). - -Finally, while Guix and Nix do allow preciesly reproducible environments, it requires extra effort. -For example simply running \inlinecode{guix install gcc} will install the most recent version of GCC that can be different at different times. -Hence, similar to the discussion in host operating system package managers, it is up to the user to ensure that their created environment is recorded properly for reproducibility in the future. -Generally, this is a major limitation of projects that rely on detached package managers for building their software, including the other tools mentioned below. - -\subsubsection{Conda/Anaconda} -\label{appendix:conda} -Conda is an independent package manager that can be used on GNU/Linux, macOS, or Windows operating systems, although all software packages are not available in all operating systems. -Conda is able to maintain an approximately independent environment on an operating system without requiring root access. - -Conda tracks the dependencies of a package/environment through a YAML formatted file, where the necessary software and their acceptable versions are listed. -However, it is not possible to fix the versions of the dependencies through the YAML files alone. -This is thoroughly discussed under issue 787 (in May 2019) of \inlinecode{conda-forge}\footnote{\url{https://github.com/conda-forge/conda-forge.github.io/issues/787}}. -In that discussion, the authors of \citeappendix{uhse19} report that the half-life of their environment (defined in a YAML file) is 3 months, and that at least one of their their dependencies breaks shortly after this period. -The main reply they got in the discussion is to build the Conda environment in a container, which is also the suggested solution by \citeappendix{gruning18}. -However, as described in Appendix \ref{appendix:independentenvironment} containers just hide the reproducibility problem, they do not fix it: containers are not static and need to evolve (i.e., re-built) with the project. -Given these limitations, \citeappendix{uhse19} are forced to host their conda-packaged software as tarballs on a separate repository. - -Conda installs with a shell script that contains a binary-blob (+500 megabytes, embedded in the shell script). -This is the first major issue with Conda: from the shell script, it is not clear what is in this binary blob and what it does. -After installing Conda in any location, users can easily activate that environment by loading a special shell script into their shell. -However, the resulting environment is not fully independent of the host operating system as described below: - -\begin{itemize} -\item The Conda installation directory is present at the start of environment variables like \inlinecode{PATH} (which is used to find programs to run) and other such environment variables. - However, the host operating system's directories are also appended afterwards. - Therefore, a user, or script may not notice that a software that is being used is actually coming from the operating system, not the controlled Conda installation. - -\item Generally, by default Conda relies heavily on the operating system and does not include core analysis components like \inlinecode{mkdir}, \inlinecode{ls} or \inlinecode{cp}. - Although they are generally the same between different Unix-like operating systems, they have their differences. - For example \inlinecode{mkdir -p} is a common way to build directories, but this option is only available with GNU Coreutils (default on GNU/Linux systems). - Running the same command within a Conda environment on a macOS for example, will crash. - Important packages like GNU Coreutils are available in channels like conda-forge, but they are not the default. - Therefore, many users may not recognize this, and failing to account for it, will cause unexpected crashes. - -\item Many major Conda packaging ``channels'' (for example the core Anaconda channel, or very popular conda-forge channel) do not include the C library, that a package was built with, as a dependency. - They rely on the host operating system's C library. - C is the core language of modern operating systems and even higher-level languages like Python or R are written in it, and need it to run. - Therefore if the host operating system's C library is different from the C library that a package was built with, a Conda-packaged program will crash and the project will not be executable. - Theoretically, it is possible to define a new Conda ``channel'' which includes the C library as a dependency of its software packages, but it will take too much time for any individual team to practically implement all their necessary packages, up to their high-level science software. - -\item Conda does allow a package to depend on a special build of its prerequisites (specified by a checksum, fixing its version and the version of its dependencies). - However, this is rarely practiced in the main Git repositories of channels like Anaconda and conda-forge: only the name of the high-level prerequisite packages is listed in a package's \inlinecode{meta.yaml} file, which is version-controlled. - Therefore two builds of the package from the same Git repository will result in different tarballs (depending on what prerequisites were present at build time). - In the Conda tarball (that contains the binaries and is not under version control) \inlinecode{meta.yaml} does include the exact versions of most build-time dependencies. - However, because the different software of one project may have been built at different times, if they depend on different versions of a single software there will be a conflict and the tarball cannot be rebuilt, or the project cannot be run. -\end{itemize} - -As reviewed above, the low-level dependence of Conda on the host operating system's components and build-time conditions, is the primary reason that it is very fast to install (thus making it an attractive tool to software developers who just need to reproduce a bug in a few minutes). -However, these same factors are major caveats in a scientific scenario, where long-term archivability, readability or usability are important. % alternative to `archivability`? - - -\subsubsection{Spack} -Spack is a package manager that is also influenced by Nix (similar to GNU Guix), see \citeappendix{gamblin15}. - But unlike Nix or GNU Guix, it does not aim for full, bit-wise reproducibility and can be built without root access in any generic location. - It relies on the host operating system for the C library. - - Spack is fully written in Python, where each software package is an instance of a class, which defines how it should be downloaded, configured, built and installed. - Therefore if the proper version of Python is not present, Spack cannot be used and when incompatibilities arise in future versions of Python (similar to how Python 3 is not compatible with Python 2), software building recipes, or the whole system, have to be upgraded. - Because of such bootstrapping problems (for example how Spack needs Python to build Python and other software), it is generally a good practice to use simpler, lower-level languages/systems for a low-level operation like package management. - - -In conclusion for all package managers, there are two common issues regarding generic package managers that hinders their usage for high-level scientific projects, as listed below: -\begin{itemize} -\item {\bf\small Pre-compiled/binary downloads:} Most package managers (excluding Nix or its derivatives) only download the software in a binary (pre-compiled) format. - This allows users to download it very fast and almost instantaneously be able to run it. - However, to provide for this, servers need to keep binary files for each build of the software on different operating systems (for example Conda needs to keep binaries for Windows, macOS and GNU/Linux operating systems). - It is also necessary for them to store binaries for each build, which includes different versions of its dependencies. - Maintaining such a large binary library is expensive, therefore once the shelf-life of a binary has expired, it will be removed, causing problems for projects that depends on them. - -\item {\bf\small Adding high-level software:} Packaging new software is not trivial and needs a good level of knowledge/experience with that package manager. -For example each has its own special syntax/standards/languages, with pre-defined variables that must already be known before someone can packaging new software for them. - -However, in many research projects, the most high-level analysis software are written by the team that is doing the research, and they are its primary users, even when the software are distributed with free licenses on open repositories. -Although active package manager members are commonly very supportive in helping to package new software, many teams may not be able to make that extra effort/time investment. -As a result, they manually install their high-level software in an uncontrolled, or non-standard way, thus jeopardizing the reproducibility of the whole work. -This is another consequence of detachment of the package manager from the project doing the analysis. -\end{itemize} - -Addressing these issues has been the basic reason d'\^etre of the proposed criteria: based on the completeness criteria, instructions to download and build the packages are included within the actual science project and no special/new syntax/language is used: software download, building and installation is done with the same language/syntax that researchers manage their research: using the shell (by default GNU Bash in Maneage) and Make (by default, GNU Make in Maneage). - - - -\subsection{Version control} -\label{appendix:versioncontrol} -A scientific project is not written in a day; it usually takes more than a year. -During this time, the project evolves significantly from its first starting date and components are added or updated constantly as it approaches completion. -Added with the complexity of modern computational projects, is not trivial to manually track this evolution, and the evolution's affect of on the final output: files produced in one stage of the project can mistakenly be used by an evolved analysis environment in later stages (where the project has evolved). - -Furthermore, scientific projects do not progress linearly: earlier stages of the analysis are often modified after later stages are written. -This is a natural consequence of the scientific method; where progress is defined by experimentation and modification of hypotheses (results from earlier phases). - -It is thus very important for the integrity of a scientific project that the state/version of its processing is recorded as the project evolves. -For example better methods are found or more data arrive. -Any intermediate dataset that is produced should also be tagged with the version of the project at the time it was created. -In this way, later processing stages can make sure that they can safely be used, i.e., no change has been made in their processing steps. - -Solutions to keep track of a project's history have existed since the early days of software engineering in the 1970s and they have constantly improved over the last decades. -Today the distributed model of ``version control'' is the most common, where the full history of the project is stored locally on different systems and can easily be integrated. -There are many existing version control solutions, for example CVS, SVN, Mercurial, GNU Bazaar, or GNU Arch. -However, currently, Git is by far the most commonly used in individual projects. -Git is also the foundation on which this paper's proof of concept (Maneage) is built upon. -Archival systems aiming for long term preservation of software like Software Heritage \citeappendix{dicosmo18} are also modeled on Git. -Hence we will just review Git here, but the general concept of version control is the same in all implementations. - -\subsubsection{Git} -With Git, changes in a project's contents are accurately identified by comparing them with their previous version in the archived Git repository. -When the user decides the changes are significant compared to the archived state, they can ``commit'' the changes into the history/repository. -The commit involves copying the changed files into the repository and calculating a 40 character checksum/hash that is calculated from the files, an accompanying ``message'' (a narrative description of the purpose/goals of the changes), and the previous commit (thus creating a ``chain'' of commits that are strongly connected to each other like Figure \ref{fig:branching}). -For example \inlinecode{f4953cc\-f1ca8a\-33616ad\-602ddf\-4cd189\-c2eff97b} is a commit identifier in the Git history of this project. -Commits are is commonly summarized by the checksum's first few characters, for example \inlinecode{f4953cc}. - -With Git, making parallel ``branches'' (in the project's history) is very easy and its distributed nature greatly helps in the parallel development of a project by a team. -The team can host the Git history on a webpage and collaborate through that. -There are several Git hosting services for example \href{http://codeberg.org}{codeberg.org}, \href{http://gitlab.com}{gitlab.com}, \href{http://bitbucket.org}{bitbucket.org} or \href{http://github.com}{github.com} (among many others). -Storing the changes in binary files is also possible in Git, however it is most useful for human-readable plain-text sources. - - - -\subsection{Job management} -\label{appendix:jobmanagement} -Any analysis will involve more than one logical step. -For example it is first necessary to download a dataset and do some preparations on it before applying the research software on it, and finally to make visualizations/tables that can be imported into the final report. -Each one of these is a logically independent step, which needs to be run before/after the others in a specific order. - -Hence job management is a critical component of a research project. -There are many tools for managing the sequence of jobs, below we review the most common ones that are also used the existing reproducibility solutions of Appendix \ref{appendix:existingsolutions}. - -\subsubsection{Manual operation with narrative} -\label{appendix:manual} -The most commonly used workflow system for many researchers is to run the commands, experiment on them and keep the output when they are happy with it. -As an improvement, some also keep a narrative description of what they ran. -Atleast in our personal experience with colleagues, this method is still being heavily practiced by many researchers. -Given that many researchers do not get trained well in computational methods, this is not surprizing and as discussed in Section \ref{discussion}, we believe that improved literacy in computational methods is the single most important factor for the integrity/reproducibility of modern science. - -\subsubsection{Scripts} -\label{appendix:scripts} -Scripts (in any language, for example GNU Bash, or Python) are the most common ways of organizing a series of steps. -They are primarily designed to execute each step sequentially (one after another), making them also very intuitive. -However, as the series of operations become complex and large, managing the workflow in a script will become highly complex. - -For example if 90\% of a long project is already done and a researcher wants to add a followup step, a script will go through all the previous steps (which can take significant time). -In other scenarios, when a small step in the middle of an analysis has to be changed, the full analysis needs to be re-run from the start. -Scripts have no concept of dependencies, forcing authors to ``temporarily'' comment parts of that they do not want to be re-run (forgetting to un-comment such parts are the most common cause of frustration for the authors and others attempting to reproduce the result). - -Such factors discourage experimentation, which is a critical component of the scientific method. -It is possible to manually add conditionals all over the script to add dependencies or only run certain steps at certain times, but they just make it harder to read, and introduce many bugs themselves. -Parallelization is another drawback of using scripts. -While its not impossible, because of the high-level nature of scripts, it is not trivial and parallelization can also be very inefficient or buggy. - - -\subsubsection{Make} -\label{appendix:make} -Make was originally designed to address the problems mentioned above for scripts \citeappendix{feldman79}. -In particular, it addresses the context of managing the compilation of software programs that involve many source code files. -With Make, the source files of a program that have not been changed are not recompiled. -Moreover, when two source files do not depend on each other, and both need to be rebuilt, they can be built in parallel. -This was found to greatly help in debugging software projects, and in speeding up test builds, giving Make a core place in software development over the last 40 years. - -The most common implementation of Make, since the early 1990s, is GNU Make. -Make was also the framework used in the first attempts at reproducible scientific papers \cite{claerbout1992,schwab2000}. -Our proof-of-concept (Maneage) also uses Make to organize its workflow. -Here, we complement that section with more technical details on Make. - -Usually, the top-level Make instructions are placed in a file called Makefile, but it is also common to use the \inlinecode{.mk} suffix for custom file names. -Each stage/step in the analysis is defined through a \emph{rule}. -Rules define \emph{recipes} to build \emph{targets} from \emph{pre-requisites}. -In \new{Unix-like operating systems}, everything is a file, even directories and devices. -Therefore all three components in a rule must be files on the running filesystem. - -To decide which operation should be re-done when executed, Make compares the time stamp of the targets and prerequisites. -When any of the prerequisite(s) is newer than a target, the recipe is re-run to re-build the target. -When all the prerequisites are older than the target, that target does not need to be rebuilt. -The recipe can contain any number of commands, they should just all start with a \inlinecode{TAB}. -Going deeper into the syntax of Make is beyond the scope of this paper, but we recommend interested readers to consult the GNU Make manual for a nice introduction\footnote{\inlinecode{\url{http://www.gnu.org/software/make/manual/make.pdf}}}. - -\subsubsection{Snakemake} -is a Python-based workflow management system, inspired by GNU Make (which is the job organizer in Maneage), that is aimed at reproducible and scalable data analysis \citeappendix{koster12}\footnote{\inlinecode{\url{https://snakemake.readthedocs.io/en/stable}}}. -It defines its own language to implement the ``rule'' concept in Make within Python. -Currently it requires Python 3.5 (released in September 2015) and above, while Snakemake was originally introduced in 2012. -Hence it is not clear if older Snakemake source files can be executed today. -This as reviewed in many tools here, this is a major longevity problem when using highlevel tools as the skeleton of the workflow. -Technically, calling commond-line programs within Python is very slow and using complex shell scripts in each step will involve a lot quotations that make the code hard to read. - -\subsubsection{Bazel} -Bazel\footnote{\inlinecode{\url{https://bazel.build}}} is a high-level job organizer that depends on Java and Python and is primarily tailored to software developers (with features like facilitating linking of libraries through its high level constructs). - -\subsubsection{SCons} -\label{appendix:scons} -Scons is a Python package for managing operations outside of Python (in contrast to CGAT-core, discussed below, which only organizes Python functions). -In many aspects it is similar to Make, for example it is managed through a `SConstruct' file. -Like a Makefile, SConstruct is also declarative: the running order is not necessarily the top-to-bottom order of the written operations within the file (unlike the imperative paradigm which is common in languages like C, Python, or FORTRAN). -However, unlike Make, SCons does not use the file modification date to decide if it should be remade. -SCons keeps the MD5 hash of all the files (in a hidden binary file) to check if the contents have changed. - -SCons thus attempts to work on a declarative file with an imperative language (Python). -It also goes beyond raw job management and attempts to extract information from within the files (for example to identify the libraries that must be linked while compiling a program). -SCons is therefore more complex than Make and its manual is almost double that of GNU Make. -Besides added complexity, all these ``smart'' features decrease its performance, especially as files get larger and more numerous: on every call, every file's checksum has to be calculated, and a Python system call has to be made (which is computationally expensive). - -Finally, it has the same drawback as any other tool that uses high-level languages, see Section \ref{appendix:highlevelinworkflow}. -We encountered such a problem while testing SCons: on the Debian-10 testing system, the \inlinecode{python} program pointed to Python 2. -However, since Python 2 is now obsolete, SCons was built with Python 3 and our first run crashed. -To fix it, we had to either manually change the core operating system path, or the SCons source hashbang. -The former will conflict with other system tools that assume \inlinecode{python} points to Python-2, the latter may need root permissions for some systems. -This can also be problematic when a Python analysis library, may require a Python version that conflicts with the running SCons. - -\subsubsection{CGAT-core} -CGAT-Core is a Python package for managing workflows, see \citeappendix{cribbs19}. -It wraps analysis steps in Python functions and uses Python decorators to track the dependencies between tasks. -It is used papers like \citeappendix{jones19}, but as mentioned in \citeappendix{jones19} it is good for managing individual outputs (for example separate figures/tables in the paper, when they are fully created within Python). -Because it is primarily designed for Python tasks, managing a full workflow (which includes many more components, written in other languages) is not trivial in it. -Another drawback with this workflow manager is that Python is a very high-level language where future versions of the language may no longer be compatible with Python 3, that CGAT-core is implemented in (similar to how Python 2 programs are not compatible with Python 3). - -\subsubsection{Guix Workflow Language (GWL)} -GWL is based on the declarative language that GNU Guix uses for package management (see Appendix \ref{appendix:packagemanagement}), which is itself based on the general purpose Scheme language. -It is closely linked with GNU Guix and can even install the necessary software needed for each individual process. -Hence in the GWL paradigm, software installation and usage does not have to be separated. -GWL has two high-level concepts called ``processes'' and ``workflows'' where the latter defines how multiple processes should be executed together. - -In conclusion, shell scripts and Make are very common and extensively used by users of Unix-based OSs (which are most commonly used for computations). -They have also existed for several decades and are robust and mature. -Many researchers are also already familiar with them and have already used them. -As we see in this appendix, the list of necessary tools for the various stages of a research project (an independent environment, package managers, job organizers, analysis languages, writing formats, editors and etc) is already very large. -Each software has its own learning curve, which is a heavy burden for a natural or social scientist for example. -Most other workflow management tools are yet another language that have to be mastered. - -Furthermore, high-level and specific solutions will evolve very fast causing disruptions in the reproducible framework. -A good example is Popper \citeappendix{jimenez17} which initially organized its workflow through the HashiCorp configuration language (HCL) because it was the default in GitHub. -However, in September 2019, GitHub dropped HCL as its default configuration language, so Popper is now using its own custom YAML-based workflow language, see Appendix \ref{appendix:popper} for more on Popper. - -\subsubsection{Nextflow (2013)} -Nextflow\footnote{\inlinecode{\url{https://www.nextflow.io}}} \citeappendix{tommaso17} workflow language with a command-line interface that is written in Java. - -\subsubsection{Generic workflow specifications (CWL and WDL)} -Due to the variety of custom workflows used in existing reproducibility solution (like those of Appendix \ref{appendix:existingsolutions}), some attempts have been made to define common workflow standards like the Common workflow language (CWL\footnote{\inlinecode{\url{https://www.commonwl.org}}}, with roots in Make, formatted in YAML or JSON) and Workflow Description Language (WDL\footnote{\inlinecode{\url{https://openwdl.org}}}, formatted in JSON). -These are primarily specifications/standards rather than software, so ideally translators can be written between the various workflow systems to make them more interoperable. - - -\subsection{Editing steps and viewing results} -\label{appendix:editors} -In order to later reproduce a project, the analysis steps must be stored in files. -For example Shell, Python or R scripts, Makefiles, Dockerfiles, or even the source files of compiled languages like C or FORTRAN. -Given that a scientific project does not evolve linearly and many edits are needed as it evolves, it is important to be able to actively test the analysis steps while writing the project's source files. -Here we review some common methods that are currently used. - -\subsubsection{Text editors} -The most basic way to edit text files is through simple text editors which just allow viewing and editing such files, for example \inlinecode{gedit} on the GNOME graphic user interface. -However, working with simple plain text editors like \inlinecode{gedit} can be very frustrating since its necessary to save the file, then go to a terminal emulator and execute the source files. -To solve this problem there are advanced text editors like GNU Emacs that allow direct execution of the script, or access to a terminal within the text editor. -However, editors that can execute or debug the source (like GNU Emacs), just run external programs for these jobs (for example GNU GCC, or GNU GDB), just as if those programs was called from outside the editor. - -With text editors, the final edited file is independent of the actual editor and can be further edited with another editor, or executed without it. -This is a very important feature that is not commonly present for other solutions mentioned below. -Another very important advantage of advanced text editors like GNU Emacs or Vi(m) is that they can also be run without a graphic user interface, directly on the command-line. -This feature is critical when working on remote systems, in particular high performance computing (HPC) facilities that do not provide a graphic user interface. -Also, the commonly used minimalistic containers do not include a graphic user interface. - -\subsubsection{Integrated Development Environments (IDEs)} -To facilitate the development of source files, IDEs add software building and running environments as well as debugging tools to a plain text editor. -Many IDEs have their own compilers and debuggers, hence source files that are maintained in IDEs are not necessarily usable/portable on other systems. -Furthermore, they usually require a graphic user interface to run. -In summary IDEs are generally very specialized tools, for special projects and are not a good solution when portability (the ability to run on different systems and at different times) is required. - -\subsubsection{Jupyter} -\label{appendix:jupyter} -Jupyter (initially IPython) \citeappendix{kluyver16} is an implementation of Literate Programming \citeappendix{knuth84}. -The main user interface is a web-based ``notebook'' that contains blobs of executable code and narrative. -Jupyter uses the custom built \inlinecode{.ipynb} format\footnote{\inlinecode{\url{https://nbformat.readthedocs.io/en/latest}}}. -Jupyter's name is a combination of the three main languages it was designed for: Julia, Python and R. -The \inlinecode{.ipynb} format, is a simple, human-readable (can be opened in a plain-text editor) file, formatted in JavaScript Object Notation (JSON). -It contains various kinds of ``cells'', or blobs, that can contain narrative description, code, or multi-media visualizations (for example images/plots), that are all stored in one file. -The cells can have any order, allowing the creation of a literal programming style graphical implementation, where narrative descriptions and executable patches of code can be intertwined. -For example to have a paragraph of text about a patch of code, and run that patch immediately in the same page. - -The \inlinecode{.ipynb} format does theoretically allow dependency tracking between cells, see IPython mailing list (discussion started by Gabriel Becker from July 2013\footnote{\url{https://mail.python.org/pipermail/ipython-dev/2013-July/010725.html}}). -Defining dependencies between the cells can allow non-linear execution which is critical for large scale (thousands of files) and complex (many dependencies between the cells) operations. -It allows automation, run-time optimization (deciding not to run a cell if its not necessary) and parallelization. -However, Jupyter currently only supports a linear run of the cells: always from the start to the end. -It is possible to manually execute only one cell, but the previous/next cells that may depend on it, also have to be manually run (a common source of human error, and frustration for complex operations). -Integration of directional graph features (dependencies between the cells) into Jupyter has been discussed, but as of this publication, there is no plan to implement it (see Jupyter's GitHub issue 1175\footnote{\inlinecode{\url{https://github.com/jupyter/notebook/issues/1175}}}). - -The fact that the \inlinecode{.ipynb} format stores narrative text, code and multi-media visualization of the outputs in one file, is another major hurdle: -The files can easy become very large (in volume/bytes) and hard to read. -Both are critical for scientific processing, especially the latter: when a web browser with proper JavaScript features is not available (can happen in a few years). -This is further exacerbated by the fact that binary data (for example images) are not directly supported in JSON and have to be converted into much less memory-efficient textual encodings. - -Finally, Jupyter has an extremely complex dependency graph: on a clean Debian 10 system, Pip (a Python package manager that is necessary for installing Jupyter) required 19 dependencies to install, and installing Jupyter within Pip needed 41 dependencies! -\citeappendix{hinsen15} reported such conflicts when building Jupyter into the Active Papers framework (see Appendix \ref{appendix:activepapers}). -However, the dependencies above are only on the server-side. -Since Jupyter is a web-based system, it requires many dependencies on the viewing/running browser also (for example special JavaScript or HTML5 features, which evolve very fast). -As discussed in Appendix \ref{appendix:highlevelinworkflow} having so many dependencies is a major caveat for any system regarding scientific/long-term reproducibility (as opposed to industrial/immediate reproducibility). -In summary, Jupyter is most useful in manual, interactive and graphical operations for temporary operations (for example educational tutorials). - - - - - - -\subsection{Project management in high-level languages} -\label{appendix:highlevelinworkflow} - -Currently the most popular high-level data analysis language is Python. -R is closely tracking it, and has superseded Python in some fields, while Julia \citeappendix{bezanson17} is quickly gaining ground. -These languages have themselves superseded previously popular languages for data analysis of the previous decades, for example Java, Perl or C++. -All are part of the C-family programming languages. -In many cases, this means that the tools to use that language are written in C, which is the language of modern operating systems. - -Scientists, or data analysts, mostly use these higher-level languages. -Therefore they are naturally drawn to also apply the higher-level languages for lower-level project management, or designing the various stages of their workflow. -For example Conda or Spack (Appendix \ref{appendix:packagemanagement}), CGAT-core (Appendix \ref{appendix:jobmanagement}), Jupyter (Appendix \ref{appendix:editors}) or Popper (Appendix \ref{appendix:popper}) are written in Python. -The discussion below applies to both the actual analysis software and project management software. -In this context, its more focused on the latter. - -Because of their nature, higher-level languages evolve very fast, creating incompatibilities on the way. -The most prominent example is the transition from Python 2 (released in 2000) to Python 3 (released in 2008). -Python 3 was incompatible with Python 2 and it was decided to abandon the former by 2015. -However, due to community pressure, this was delayed to January 1st, 2020. -The end-of-life of Python 2 caused many problems for projects that had invested heavily in Python 2: all their previous work had to be translated, for example see \citeappendix{jenness17} or Appendix \ref{appendix:sciunit}. -Some projects could not make this investment and their developers decided to stop maintaining it, for example VisTrails (see Appendix \ref{appendix:vistrails}). - -The problems were not just limited to translation. -Python 2 was still actively being actively used during the transition period (and is still being used by some, after its end-of-life). -Therefore, developers of packages used by others had to maintain (for example fix bugs in) both versions in one package. -This is not particular to Python, a similar evolution occurred in Perl: in 2000 it was decided to improve Perl 5, but the proposed Perl 6 was incompatible with it. -However, the Perl community decided not to abandon Perl 5, and Perl 6 was eventually defined as a new language that is now officially called ``Raku'' (\url{https://raku.org}). - -It is unreasonably optimistic to assume that high-level languages will not undergo similar incompatible evolutions in the (not too distant) future. -For software developers, this is not a problem at all: non-scientific software, and the general population's usage of them, has a similarly fast evolvution. -Hence, it is rarely (if ever) necessary to look into codes that are more than a couple of years old. -However, in the sciences (which are commonly funded by public money) this is a major caveat for the longer-term usability of solutions that are designed in such high level languages. - -In summary, in this section we are discussing the bootstrapping problem as regards scientific projects: the workflow/pipeline can reproduce the analysis and its dependencies, but the dependencies of the workflow itself cannot not be ignored. -Beyond technical, low-level, problems for the developers mentioned above, this causes major problems for scientific project management as listed below: - -\subsubsection{Dependency hell} -The evolution of high-level languages is extremely fast, even within one version. -For example packages that are written in Python 3 often only work with a special interval of Python 3 versions (for example newer than Python 3.6). -This is not just limited to the core language, much faster changes occur in their higher-level libraries. -For example version 1.9 of Numpy (Python's numerical analysis module) discontinued support for Numpy's predecessor (called Numeric), causing many problems for scientific users \citeappendix{hinsen15}. - -On the other hand, the dependency graph of tools written in high-level languages is often extremely complex. -For example see Figure 1 of \cite{alliez19}, it shows the dependencies and their inter-dependencies for Matplotlib (a popular plotting module in Python). -Acceptable version intervals between the dependencies will cause incompatibilities in a year or two, when a robust package manager is not used (see Appendix \ref{appendix:packagemanagement}). - -Since a domain scientist does not always have the resources/knowledge to modify the conflicting part(s), many are forced to create complex environments with different versions of Python and pass the data between them (for example just to use the work of a previous PhD student in the team). -This greatly increases the complexity of the project, even for the principal author. -A good reproducible workflow can account for these different versions. -However, when the actual workflow system (not the analysis software) is written in a high-level language this will cause a major problem. - -For example, merely installing the Python installer (\inlinecode{pip}) on a Debian system (with \inlinecode{apt install pip2} for Python 2 packages), required 32 other packages as dependencies. -\inlinecode{pip} is necessary to install Popper and Sciunit (Appendices \ref{appendix:popper} and \ref{appendix:sciunit}). -As of this writing, the \inlinecode{pip3 install popper} and \inlinecode{pip2 install sciunit2} commands for installing each, required 17 and 26 Python modules as dependencies. -It is impossible to run either of these solutions if there is a single conflict in this very complex dependency graph. -This problem actually occurred while we were testing Sciunit: even though it installed, it could not run because of conflicts (its last commit was only 1.5 years old), for more see Appendix \ref{appendix:sciunit}. -\citeappendix{hinsen15} also report a similar problem when attempting to install Jupyter (see Appendix \ref{appendix:editors}). -Of course, this also applies to tools that these systems use, for example Conda (which is also written in Python, see Appendix \ref{appendix:packagemanagement}). - - - - - -\subsubsection{Generational gap} -This occurs primarily for domain scientists (for example astronomers, biologists or social sciences). -Once they have mastered one version of a language (mostly in the early stages of their career), they tend to ignore newer versions/languages. -The inertia of programming languages is very strong. -This is natural, because they have their own science field to focus on, and re-writing their high-level analysis toolkits (which they have curated over their career and is often only readable/usable by themselves) in newer languages every few years requires too much investment and time. - -When this investment is not possible, either the mentee has to use the mentor's old method (and miss out on all the new tools, which they need for the future job prospects), or the mentor has to avoid implementation details in discussions with the mentee, because they do not share a common language. -The authors of this paper have personal experiences in both mentor/mentee relational scenarios. -This failure to communicate in the details is a very serious problem, leading to the loss of valuable inter-generational experience. - - - - - - - - - - - - - - - - - - - - -\section{Survey of common existing reproducible workflows} -\label{appendix:existingsolutions} - -As reviewed in the introduction, the problem of reproducibility has received a lot of attention over the last three decades and various solutions have already been proposed. -The core principles that many of the existing solutions (including Maneage) aim to achieve are nicely summarized by the FAIR principles \citeappendix{wilkinson16}. -In this appendix, some of the solutions are reviewed. -The solutions are based on an evolving software landscape, therefore they are ordered by date: when the project has a webpage, the year of its first release is used for the sorting, otherwise their paper's publication year is used. - -For each solution, we summarize its methodology and discuss how it relates to the criteria proposed in this paper. -Freedom of the software/method is a core concept behind scientific reproducibility, as opposed to industrial reproducibility where a black box is acceptable/desirable. -Therefore proprietary solutions like Code Ocean\footnote{\inlinecode{\url{https://codeocean.com}}} or Nextjournal\footnote{\inlinecode{\url{https://nextjournal.com}}} will not be reviewed here. -Other studies have also attempted to review existing reproducible solutions, foro example \citeappendix{konkol20}. - -\subsection{Suggested rules, checklists, or criteria} -Before going into the various implementations, it is also useful to review existing suggested rules, checklists or criteria for computationally reproducible research. - -All the cases below are primarily targetted to immediate reproducibility and do not consider longevity explicitly. -Therefore, they lack a strong/clear completeness criterion (they mainly only suggest, rather than require, the recording of versions, and their ultimate suggestion of storing the full binary OS in a binary VM or container is problematic (as mentioned in \ref{appendix:independentenvironment} and \citeappendix{oliveira18}). - -Sandve et al. \citeappendix{sandve13} propose ``ten simple rules for reproducible computational research'' that can be applied in any project. -Generally, these are very similar to the criteria proposed here and follow a similar spirit, but they do not provide any actual research papers following up all those points, nor do they provide a proof of concept. -The Popper convention \citeappendix{jimenez17} also provides a set of principles that are indeed generally useful, among which some are common to the criteria here (for example, automatic validation, and, as in Maneage, the authors suggest providing a template for new users), -but the authors do not include completeness as a criterion nor pay attention to longevity (Popper itself is written in Python with many dependencies, and its core operating language has already changed once). -For more on Popper, please see Section \ref{appendix:popper}. - -For improved reproducibility in Jupyter notebook users, \citeappendix{rule19} propose ten rules to improve reproducibility and also provide links to example implementations. -These can be very useful for users of Jupyter, but are not generic for non-Jupyter-based computational projects. -Some criteria (which are indeed very good in a more general context) do not directly relate to reproducibility, for example their Rule 1: ``Tell a Story for an Audience''. -Generally, as reviewed in Sections \ref{sec:longevityofexisting} and \ref{appendix:jupyter}, Jupyter itself has many issues regarding reproducibility. - -To create Docker images, N\"ust et al. propose ``ten simple rules'' in \citeappendix{nust20}. -They recommend some issues that can indeed help increase the quality of Docker images and their production/usage, such as their rule 7 to ``mount datasets [only] at run time'' to separate the computational environment from the data. -However, long-term reproducibility of the images is not included as a criterion by these authors. -For example, they recommend using base operating systems, with version identification limited to a single brief identifier such as \inlinecode{ubuntu:18.04}, which has a serious problem with longevity issues (Section \ref{sec:longevityofexisting}). -Furthermore, in their proof-of-concept Dockerfile (listing 1), \inlinecode{rocker} is used with a tag (not a digest), which can be problematic due to the high risk of ambiguity (as discussed in Section \ref{appendix:containers}). - -\subsection{Reproducible Electronic Documents, RED (1992)} -\label{appendix:red} - -RED\footnote{\inlinecode{\url{http://sep.stanford.edu/doku.php?id=sep:research:reproducible}}} is the first attempt that we could find on doing reproducible research, see \cite{claerbout1992,schwab2000}. -It was developed within the Stanford Exploration Project (SEP) for Geophysics publications. -Their introductions on the importance of reproducibility, resonate a lot with today's environment in computational sciences. -In particular the heavy investment one has to make in order to re-do another scientist's work, even in the same team. -RED also influenced other early reproducible works, for example \citeappendix{buckheit1995}. - -To orchestrate the various figures/results of a project, from 1990, they used ``Cake'' \citeappendix{somogyi87}, a dialect of Make, for more on Make, see Appendix \ref{appendix:jobmanagement}. -As described in \cite{schwab2000}, in the latter half of that decade, they moved to GNU Make, which was much more commonly used, developed and came with a complete and up-to-date manual. -The basic idea behind RED's solution was to organize the analysis as independent steps, including the generation of plots, and organizing the steps through a Makefile. -This enabled all the results to be re-executed with a single command. -Several basic low-level Makefiles were included in the high-level/central Makefile. -The reader/user of a project had to manually edit the central Makefile and set the variable \inlinecode{RESDIR} (result dir), this is the directory where built files are kept. -Afterwards, the reader could set which figures/parts of the project to reproduce by manually adding its name in the central Makefile, and running Make. - -At the time, Make was already practiced by individual researchers and projects as a job orchestration tool, but SEP's innovation was to standardize it as an internal policy, and define conventions for the Makefiles to be consistent across projects. -This enabled new members to benefit from the already existing work of previous team members (who had graduated or moved to other jobs). -However, RED only used the existing software of the host system, it had no means to control them. -Therefore, with wider adoption, they confronted a ``versioning problem'' where the host's analysis software had different versions on different hosts, creating different results, or crashing \citeappendix{fomel09}. -Hence in 2006 SEP moved to a new Python-based framework called Madagascar, see Appendix \ref{appendix:madagascar}. - - - - - -\subsection{Apache Taverna (2003)} -\label{appendix:taverna} -Apache Taverna\footnote{\inlinecode{\url{https://taverna.incubator.apache.org}}} \citeappendix{oinn04} is a workflow management system written in Java with a graphical user interface which is still being developed. -A workflow is defined as a directed graph, where nodes are called ``processors''. -Each Processor transforms a set of inputs into a set of outputs and they are defined in the Scufl language (an XML-based language, were each step is an atomic task). -Other components of the workflow are ``Data links'' and ``Coordination constraints''. -The main user interface is graphical, where users move processors in the given space and define links between their inputs outputs (manually constructing a lineage like Figure \ref{fig:datalineage}). -Taverna is only a workflow manager and is not integrated with a package manager, hence the versions of the used software can be different in different runs. -\citeappendix{zhao12} have studied the problem of workflow decays in Taverna. - - - - - -\subsection{Madagascar (2003)} -\label{appendix:madagascar} -Madagascar\footnote{\inlinecode{\url{http://ahay.org}}} \citeappendix{fomel13} is a set of extensions to the SCons job management tool (reviewed in \ref{appendix:scons}). -Madagascar is a continuation of the Reproducible Electronic Documents (RED) project that was discussed in Appendix \ref{appendix:red}. -Madagascar has been used in the production of hundreds of research papers or book chapters\footnote{\inlinecode{\url{http://www.ahay.org/wiki/Reproducible_Documents}}}, 120 prior to \citeappendix{fomel13}. - -Madagascar does include project management tools in the form of SCons extensions. -However, it is not just a reproducible project management tool. -It is primarily a collection of analysis programs and tools to interact with RSF files, and plotting facilities. -The Regularly Sampled File (RSF) file format\footnote{\inlinecode{\url{http://www.ahay.org/wiki/Guide\_to\_RSF\_file\_format}}} is a custom plain-text file that points to the location of the actual data files on the filesystem and acts as the intermediary between Madagascar's analysis programs. -For example in our test of Madagascar 3.0.1, it installed 855 Madagascar-specific analysis programs (\inlinecode{PREFIX/bin/sf*}). -The analysis programs mostly target geophysical data analysis, including various project specific tools: more than half of the total built tools are under the \inlinecode{build/user} directory which includes names of Madagascar users. - -Besides the location or contents of the data, RSF also contains name/value pairs that can be used as options to Madagascar programs, which are built with inputs and outputs of this format. -Since RSF contains program options also, the inputs and outputs of Madagascar's analysis programs are read from, and written to, standard input and standard output. - -In terms of completeness, as long as the user only uses Madagascar's own analysis programs, it is fairly complete at a high level (not lower-level OS libraries). -However, this comes at the expense of a large amount of bloatware (programs that one project may never need, but is forced to build). -Also, the linking between the analysis programs (of a certain user at a certain time) and future versions of that program (that is updated in time) is not immediately obvious. -Madagascar could have been more useful to a larger community if the workflow components were maintained as a separate project compared to the analysis components. - -\subsection{GenePattern (2004)} -\label{appendix:genepattern} -GenePattern\footnote{\inlinecode{\url{https://www.genepattern.org}}} \citeappendix{reich06} (first released in 2004) is a client-server software containing many common analysis functions/modules, primarily focused for Gene studies. -Although its highly focused to a special research field, it is reviewed here because its concepts/methods are generic, and in the context of this paper. - -Its server-side software is installed with fixed software packages that are wrapped into GenePattern modules. -The modules are used through a web interface, the modern implementation is GenePattern Notebook \citeappendix{reich17}. -It is an extension of the Jupyter notebook (see Appendix \ref{appendix:editors}), which also has a special ``GenePattern'' cell that will connect to GenePattern servers for doing the analysis. -However, the wrapper modules just call an existing tool on the host system. -Given that each server may have its own set of installed software, the analysis may differ (or crash) when run on different GenePattern servers, hampering reproducibility. - -%% GenePattern shutdown announcement (although as of November 2020, it does not open any more!): https://www.genepattern.org/blog/2019/10/01/the-genomespace-project-is-ending-on-november-15-2019 -The primary GenePattern server was active since 2008 and had 40,000 registered users with 2000 to 5000 jobs running every week \citeappendix{reich17}. -However, it was shut down on November 15th 2019 due to end of funding. -All processing with this sever has stopped, and any archived data on it has been deleted. -Since GenePattern is free software, there are alternative public servers to use, so hopefully work on it will continue. -However, funding is limited and those servers may face similar funding problems. -This is a very nice example of the fragility of solutions that depend on archiving and running the research codes with high-level research products (including data, binary/compiled code which are expensive to keep) in one place. - - - - - -\subsection{Kepler (2005)} -Kepler\footnote{\inlinecode{\url{https://kepler-project.org}}} \citeappendix{ludascher05} is a Java-based Graphic User Interface workflow management tool. -Users drag-and-drop analysis components, called ``actors'', into a visual, directional graph, which is the workflow (similar to Figure \ref{fig:datalineage}). -Each actor is connected to others through the Ptolemy II\footnote{\inlinecode{\url{https://ptolemy.berkeley.edu}}} \citeappendix{eker03}. -In many aspects, the usage of Kepler and its issues for long-term reproducibility is like Apache Taverna (see Section \ref{appendix:taverna}). - - - - - -\subsection{VisTrails (2005)} -\label{appendix:vistrails} - -VisTrails\footnote{\inlinecode{\url{https://www.vistrails.org}}} \citeappendix{bavoil05} was a graphical workflow managing system. -According to its webpage, VisTrails maintainance has stopped since May 2016, its last Git commit, as of this writing, was in November 2017. -However, given that it was well maintained for over 10 years is an achievement. - -VisTrails (or ``visualization trails'') was initially designed for managing visualizations, but later grew into a generic workflow system with meta-data and provenance features. -Each analysis step, or module, is recorded in an XML schema, which defines the operations and their dependencies. -The XML attributes of each module can be used in any XML query language to find certain steps (for example those that used a certain command). -Since the main goal was visualization (as images), apparently its primary output is in the form of image spreadsheets. -Its design is based on a change-based provenance model using a custom VisTrails provenance query language (vtPQL), for more see \citeappendix{scheidegger08}. -Since XML is a plane text format, as the user inspects the data and makes changes to the analysis, the changes are recorded as ``trails'' in the project's VisTrails repository that operates very much like common version control systems (see Appendix \ref{appendix:versioncontrol}). -. -However, even though XML is in plain text, it is very hard to edit manually. -VisTrails therefore provides a graphic user interface with a visual representation of the project's inter-dependent steps (similar to Figure \ref{fig:datalineage}). -Besides the fact that it is no longer maintained, VisTrails does not control the software that is run, it only controls the sequence of steps that they are run in. - - - - - -\subsection{Galaxy (2010)} -\label{appendix:galaxy} - -Galaxy\footnote{\inlinecode{\url{https://galaxyproject.org}}} is a web-based Genomics workbench \citeappendix{goecks10}. -The main user interface are ``Galaxy Pages'', which does not require any programming: users simply use abstract ``tools'' which are a wrappers over command-line programs. -Therefore the actual running version of the program can be hard to control across different Galaxy servers. -Besides the automatically generated metadata of a project (which include version control, or its history), users can also tag/annotate each analysis step, describing its intent/purpose. -Besides some small differences Galaxy seems very similar to GenePattern (Appendix \ref{appendix:genepattern}), so most of the same points there apply here too (including the very large cost of maintining such a system). - - - - - -\subsection{Image Processing On Line journal, IPOL (2010)} -\label{appendix:ipol} -The IPOL journal\footnote{\inlinecode{\url{https://www.ipol.im}}} \citeappendix{limare11} (first published article in July 2010) publishes papers on image processing algorithms as well as the the full code of the proposed algorithm. -An IPOL paper is a traditional research paper, but with a focus on implementation. -The published narrative description of the algorithm must be detailed to a level that any specialist can implement it in their own programming language (extremely detailed). -The author's own implementation of the algorithm is also published with the paper (in C, C++ or MATLAB), the code must be commented well enough and link each part of it with the relevant part of the paper. -The authors must also submit several example datasets/scenarios. -The referee is expected to inspect the code and narrative, confirming that they match with each other, and with the stated conclusions of the published paper. -After publication, each paper also has a ``demo'' button on its webpage, allowing readers to try the algorithm on a web-interface and even provide their own input. - -IPOL has grown steadily over the last 10 years, publishing 23 research articles in 2019 alone. -We encourage the reader to visit its webpage and see some of its recent papers and their demos. -The reason it can be so thorough and complete is its very narrow scope (low-level image processing algorithms), where the published algorithms are highly atomic, not needing significant dependencies (beyond input/output of well-known formats), allowing the referees and readers to go deeply into each implemented algorithm. -In fact, high-level languages like Perl, Python or Java are not acceptable in IPOL precisely because of the additional complexities, such as dependencies, that they require. -However, many data-intensive projects commonly involve dozens of high-level dependencies, with large and complex data formats and analysis, so this solution is not scalable. - -IPOL thus fails on our Scalability criteria. -Furthermore, by not publishing/archiving each paper's version control history or directly linking the analysis and produced paper, it fails Criterias 6 and 7. -Note that on the webpage, it is possible to change parameters, but that will not affect the produced PDF. -A paper written in Maneage (the proof-of-concept solution presented in this paper) could be scrutinised at a similar detailed level to IPOL, but for much more complex research scenarios, involving hundreds of dependencies and complex processing of the data. - - - - - - -\subsection{WINGS (2010)} -\label{appendix:wings} - -WINGS\footnote{\inlinecode{\url{https://wings-workflows.org}}} \citeappendix{gil10} is an automatic workflow generation algorithm. -It runs on a centralized web server, requiring many dependencies (such that it is recommended to download Docker images). -It allows users to define various workflow components (for example datasets, analysis components and etc), with high-level goals. -It then uses selection and rejection algorithms to find the best components using a pool of analysis components that can satisfy the requested high-level constraints. -\tonote{Read more about this} - - - - - -\subsection{Active Papers (2011)} -\label{appendix:activepapers} -Active Papers\footnote{\inlinecode{\url{http://www.activepapers.org}}} attempts to package the code and data of a project into one file (in HDF5 format). -It was initially written in Java because its compiled byte-code outputs in JVM are portable on any machine \citeappendix{hinsen11}. -However, Java is not a commonly used platform today, hence it was later implemented in Python \citeappendix{hinsen15}. - -In the Python version, all processing steps and input data (or references to them) are stored in a HDF5 file. -However, it can only account for pure-Python packages using the host operating system's Python modules \tonote{confirm this!}. -When the Python module contains a component written in other languages (mostly C or C++), it needs to be an external dependency to the Active Paper. - -As mentioned in \citeappendix{hinsen15}, the fact that it relies on HDF5 is a caveat of Active Papers, because many tools are necessary to merely open it. -Downloading the pre-built HDF View binaries (provided by the HDF group) is not possible anonymously/automatically (login is required). -Installing it using the Debian or Arch Linux package managers also failed due to dependencies in our trials. -Furthermore, as a high-level data format HDF5 evolves very fast, for example HDF5 1.12.0 (February 29th, 2020) is not usable with older libraries provided by the HDF5 team. % maybe replace with: February 29\textsuperscript{th}, 2020? - -While data and code are indeed fundamentally similar concepts technically\citeappendix{hinsen16}, they are used by humans differently due to their volume: the code of a large project involving Terabytes of data can be less than a megabyte. -Hence, storing code and data together becomes a burden when large datasets are used, this was also acknowledged in \citeappendix{hinsen15}. -Also, if the data are proprietary (for example medical patient data), the data must not be released, but the methods that were applied to them can be published. -Furthermore, since all reading and writing is done in the HDF5 file, it can easily bloat the file to very large sizes due to temporary/reproducible files, and its necessary to remove/dummify them, thus complicating the code, making it hard to read. -For example the Active Papers HDF5 file of \citeappendix[in \href{https://doi.org/10.5281/zenodo.2549987}{zenodo.2549987}]{kneller19} is 1.8 giga-bytes. - -In many scenarios, peers just want to inspect the processing by reading the code and checking a very special part of it (one or two lines, just to see the option values to one step for example). -They do not necessarily need to run it, or obtaining the datasets. -Hence the extra volume for data, and obscure HDF5 format that needs special tools for reading plain text code is a major hinderance. - - - - - -\subsection{Collage Authoring Environment (2011)} -\label{appendix:collage} -The Collage Authoring Environment \citeappendix{nowakowski11} was the winner of Elsevier Executable Paper Grand Challenge \citeappendix{gabriel11}. -It is based on the GridSpace2\footnote{\inlinecode{\url{http://dice.cyfronet.pl}}} distributed computing environment\tonote{find citation}, which has a web-based graphic user interface. -Through its web-based interface, viewers of a paper can actively experiment with the parameters of a published paper's displayed outputs (for example figures). -\tonote{See how it containerizes the software environment} - - - - - -\subsection{SHARE (2011)} -\label{appendix:SHARE} -SHARE\footnote{\inlinecode{\url{https://is.ieis.tue.nl/staff/pvgorp/share}}} \citeappendix{vangorp11} is a web portal that hosts virtual machines (VMs) for storing the environment of a research project. -The top project webpage above is still active, however, the virtual machines and SHARE system have been removed since 2019, probably due to the large volume and high maintainence cost of the VMs. - -SHARE was recognized as second position in the Elsevier Executable Paper Grand Challenge \citeappendix{gabriel11}. -Simply put, SHARE is just a VM that users can download and run. -The limitations of VMs for reproducibility were discussed in Appendix \ref{appendix:virtualmachines}, and the SHARE system does not specify any requirements on making the VM itself reproducible. - - - - - -\subsection{Verifiable Computational Result, VCR (2011)} -\label{appendix:verifiableidentifier} -A ``verifiable computational result''\footnote{\inlinecode{\url{http://vcr.stanford.edu}}} is an output (table, figure, or etc) that is associated with a ``verifiable result identifier'' (VRI), see \citeappendix{gavish11}. -It was awarded the third prize in the Elsevier Executable Paper Grand Challenge \citeappendix{gabriel11}. - -A VRI is created using tags within the programming source that produced that output, also recording its version control or history. -This enables exact identification and citation of results. -The VRIs are automatically generated web-URLs that link to public VCR repositories containing the data, inputs and scripts, that may be re-executed. -According to \citeappendix{gavish11}, the VRI generation routine has been implemented in MATLAB, R and Python, although only the MATLAB version was available during the writing of this paper. -VCR also has special \LaTeX{} macros for loading the respective VRI into the generated PDF. - -Unfortunately most parts of the webpage are not complete at the time of this writing. -The VCR webpage contains an example PDF\footnote{\inlinecode{\url{http://vcr.stanford.edu/paper.pdf}}} that is generated with this system, however, the linked VCR repository\footnote{\inlinecode{\url{http://vcr-stat.stanford.edu}}} does not exist at the time of this writing. -Finally, the date of the files in the MATLAB extension tarball are set to 2011, hinting that probably VCR has been abandoned soon after the publication of \citeappendix{gavish11}. - - - - - -\subsection{SOLE (2012)} -\label{appendix:sole} -SOLE (Science Object Linking and Embedding) defines ``science objects'' (SOs) that can be manually linked with phrases of the published paper \citeappendix{pham12,malik13}. -An SO is any code/content that is wrapped in begin/end tags with an associated type and name. -For example special commented lines in a Python, R or C program. -The SOLE command-line program parses the tagged file, generating metadata elements unique to the SO (including its URI). -SOLE also supports workflows as Galaxy tools \citeappendix{goecks10}. - -For reproducibility, \citeappendix{pham12} suggest building a SOLE-based project in a virtual machine, using any custom package manager that is hosted on a private server to obtain a usable URI. -However, as described in Appendices \ref{appendix:independentenvironment} and \ref{appendix:packagemanagement}, unless virtual machines are built with robust package managers, this is not a sustainable solution (the virtual machine itself is not reproducible). -Also, hosting a large virtual machine server with fixed IP on a hosting service like Amazon (as suggested there) for every project in perpetuity will be very expensive. -The manual/artificial definition of tags to connect parts of the paper with the analysis scripts is also a caveat due to human error and incompleteness (tags the authors may not consider important, but may be useful later). -In Maneage, instead of artificial/commented tags directly link the analysis input and outputs to the paper's text automatically. - - - -\subsection{Sumatra (2012)} -Sumatra\footnote{\inlinecode{\url{http://neuralensemble.org/sumatra}}} \citeappendix{davison12} attempts to capture the environment information of a running project. -It is written in Python and is a command-line wrapper over the analysis script. -By controlling a project at running-time, Sumatra is able to capture the environment it was run in. -The captured environment can be viewed in plain text or a web interface. -Sumatra also provides \LaTeX/Sphinx features, which will link the paper with the project's Sumatra database. -This enables researchers to use a fixed version of a project's figures in the paper, even at later times (while the project is being developed). - -The actual code that Sumatra wraps around, must itself be under version control, and it does not run if there is non-committed changes (although its not clear what happens if a commit is amended). -Since information on the environment has been captured, Sumatra is able to identify if it has changed since a previous run of the project. -Therefore Sumatra makes no attempt at storing the environment of the analysis as in Sciunit (see Appendix \ref{appendix:sciunit}), but its information. -Sumatra thus needs to know the language of the running program and is not generic. -It just captures the environment, it does not store \emph{how} that environment was built. - - - - - -\subsection{Research Object (2013)} -\label{appendix:researchobject} - -The Research object\footnote{\inlinecode{\url{http://www.researchobject.org}}} is collection of meta-data ontologies, to describe aggregation of resources, or workflows, see \citeappendix{bechhofer13} and \citeappendix{belhajjame15}. -It thus provides resources to link various workflow/analysis components (see Appendix \ref{appendix:existingtools}) into a final workflow. - -\citeappendix{bechhofer13} describes how a workflow in Apache Taverna (Appendix \ref{appendix:taverna}) can be translated into research objects. -The important thing is that the research object concept is not specific to any special workflow, it is just a metadata bundle/standard which is only as robust in reproducing the result as the running workflow. - - - - -\subsection{Sciunit (2015)} -\label{appendix:sciunit} -Sciunit\footnote{\inlinecode{\url{https://sciunit.run}}} \citeappendix{meng15} defines ``sciunit''s that keep the executed commands for an analysis and all the necessary programs and libraries that are used in those commands. -It automatically parses all the executables in the script, and copies them, and their dependency libraries (down to the C library), into the sciunit. -Because the sciunit contains all the programs and necessary libraries, its possible to run it readily on other systems that have a similar CPU architecture. -Sciunit was originally written in Python 2 (which reached its end-of-life in January 1st, 2020). -Therefore Sciunit2 is a new implementation in Python 3. - -The main issue with Sciunit's approach is that the copied binaries are just black boxes: it is not possible to see how the used binaries from the initial system were built. -This is a major problem for scientific projects: in principle (not knowing how they programs were built) and in practice (archiving a large volume sciunit for every step of an analysis requires a lot of storage space). - - - - - -\subsection{Umbrella (2015)} -Umbrella \citeappendix{meng15b} is a high-level wrapper script for isolating the environment of an analysis. -The user specifies the necessary operating system, and necessary packages and analysis steps in variuos JSON files. -Umbrella will then study the host operating system and the various necessary inputs (including data and software) through a process similar to Sciunits mentioned above to find the best environment isolator (maybe using Linux containerization, containers or VMs). -We could not find a URL to the source software of Umbrella (no source code repository is mentioned in the papers we reviewed above), but from the descriptions in \citeappendix{meng17}, it is written in Python 2.6 (which is now \new{deprecated}). - - - - - -\subsection{ReproZip (2016)} -ReproZip\footnote{\inlinecode{\url{https://www.reprozip.org}}} \citeappendix{chirigati16} is a Python package that is designed to automatically track all the necessary data files, libraries and environment variables into a single bundle. -The tracking is done at the kernel system-call level, so any file that is accessed during the running of the project is identified. -The tracked files can be packaged into a \inlinecode{.rpz} bundle that can then be unpacked into another system. - -ReproZip is therefore very good to take a ``snapshot'' of the running environment into a single file. -The bundle can become very large if large, or many datasets, are used or if the software evironment is complex (many dependencies). -Since it copies the binary software libraries, it can only be run on systems with a similar CPU architecture to the original. -Furthermore, ReproZip just copies the binary/compiled files used in a project, it has no way to know how those software were built. -As mentioned in this paper, and also \citeappendix{oliveira18} the question of ``how'' the environment was built is critical for understanding the results and simply having the binaries cannot necessarily be useful. - -For the data, it is similarly not possible to extract which data server they came from. -Hence two projects that each use a 1-terabyte dataset will need a full copy of that same 1-terabyte file in their bundle, making long term preservation extremely expensive. - - - - - -\subsection{Binder (2017)} -Binder\footnote{\inlinecode{\url{https://mybinder.org}}} is used to containerize already existing Jupyter based processing steps. -Users simply add a set of Binder-recognized configuration files to their repository and Binder will build a Docker image and install all the dependencies inside of it with Conda (the list of necessary packages comes from Conda). -One good feature of Binder is that the imported Docker image must be tagged (something like a checksum). -This will ensure that future/latest updates of the imported Docker image are not mistakenly used. -However, it does not make sure that the dockerfile used by the imported Docker image follows a similar convention also. -Binder is used by \citeappendix{jones19}. - - - - - -\subsection{Gigantum (2017)} -%% I took the date from their PiPy page, where the first version 0.1 was published in November 2016. -Gigantum\footnote{\inlinecode{\url{https://gigantum.com}}} is a client/server system, in which the client is a web-based (graphical) interface that is installed as ``Gigantum Desktop'' within a Docker image. -Gigantum uses Docker containers for an independent environment, Conda (or Pip) to install packages, Jupyter notebooks to edit and run code, and Git to store its history. -Simply put, its a high-level wrapper for combining these components. -Internally, a Gigantum project is organized as files in a directory that can be opened without their own client. -The file structure (which is under version control) includes codes, input data and output data. -As acknowledged on their own webpage, this greatly reduces the speed of Git operations, transmitting, or archiving the project. -Therefore there are size limits on the dataset/code sizes. -However, there is one directory which can be used to store files that must not be tracked. - - - - -\subsection{Popper (2017)} -\label{appendix:popper} -Popper\footnote{\inlinecode{\url{https://falsifiable.us}}} is a software implementation of the Popper Convention \citeappendix{jimenez17}. -The Popper team's own solution is through a command-line program called \inlinecode{popper}. -The \inlinecode{popper} program itself is written in Python. -However, job management was initially based on the HashiCorp configuration language (HCL) because HCL was used by ``GitHub Actions'' to manage workflows. -Moreover, from October 2019 Github changed to a custom YAML-based languguage, so Popper also deprecated HCL. -This is an important issue when low-level choices are based on service providers. - -To start a project, the \inlinecode{popper} command-line program builds a template, or ``scaffold'', which is a minimal set of files that can be run. -However, as of this writing, the scaffold is not complete: it lacks a manuscript and validation of outputs (as mentioned in the convention). -By default Popper runs in a Docker image (so root permissions are necessary and reproducible issues with Docker images have been discussed above), but Singularity is also supported. -See Appendix \ref{appendix:independentenvironment} for more on containers, and Appendix \ref{appendix:highlevelinworkflow} for using high-level languages in the workflow. - -Popper does not comply with the completeness, minimal complexity and including-narrative criteria. -Moreover, the scaffold that is provided by Popper is an output of the program that is not directly under version control. -Hence, tracking future changes in Popper and how they relate to the high-level projects that depend on it will be very hard. -In Maneage, the same \inlinecode{maneage} git branch is shared by the developers and users; any new feature or change in Maneage can thus be directly tracked with Git when the high-level project merges their branch with Maneage. - - -\subsection{Whole Tale (2017)} -\label{appendix:wholetale} - -Whole Tale\footnote{\inlinecode{\url{https://wholetale.org}}} is a web-based platform for managing a project and organizing data provenance, see \citeappendix{brinckman17} -It uses online editors like Jupyter or RStudio (see Appendix \ref{appendix:editors}) that are encapsulated in a Docker container (see Appendix \ref{appendix:independentenvironment}). - -The web-based nature of Whole Tale's approach, and its dependency on many tools (which have many dependencies themselves) is a major limitation for future reproducibility. -For example, when following their own tutorial on ``Creating a new tale'', the provided Jupyter notebook could not be executed because of a dependency problem. -This was reported to the authors as issue 113\footnote{\inlinecode{\url{https://github.com/whole-tale/wt-design-docs/issues/113}}}, but as all the second-order dependencies evolve, its not hard to envisage such dependency incompatibilities being the primary issue for older projects on Whole Tale. -Furthermore, the fact that a Tale is stored as a binary Docker container causes two important problems: -1) it requires a very large storage capacity for every project that is hosted there, making it very expensive to scale if demand expands. -2) It is not possible to see how the environment was built accurately (when the Dockerfile uses \inlinecode{apt}). -This issue with Whole Tale (and generally all other solutions that only rely on preserving a container/VM) was also mentioned in \citeappendix{oliveira18}, for more on this, please see Appendix \ref{appendix:packagemanagement}. - - - - - -\subsection{Occam (2018)} -Occam\footnote{\inlinecode{\url{https://occam.cs.pitt.edu}}} \citeappendix{oliveira18} is web-based application to preserve software and its execution. -To achieve long-term reproducibility, Occam includes its own package manager (instructions to build software and their dependencies) to be in full control of the software build instructions, similar to Maneage. -Besides Nix or Guix (which are primarily a package manager that can also do job management), Occum has been the only solution in our survey here that attempts to be complete in this aspect. - -However it is incomplete from the perspective of requirements: it works within a Docker image (that requires root permissions) and currently only runs on Debian-based, Redhat-based and Arch-based GNU/Linux operating systems that respectively use the \inlinecode{apt}, \inlinecode{pacman} or \inlinecode{yum} package managers. -It is also itself written in Python (version 3.4 or above), hence it is not clear - -Furthermore, it also violates our complexity criteria because the instructions to build the software, their versions and etc are not immediately viewable or modifable by the user. -Occam contains its own JSON database for this that should be parsed with its own custom program. -The analysis phase of Occum is also through a drag-and-drop interface (similar to Taverna, Appendix \ref{appendix:taverna}) that is a web-based graphic user interface. -All the connections between various phases of an analysis need to be pre-defined in a JSON file and manually linked in the GUI. -Hence for complex data analysis operations with involve thousands of steps, it is not scalable. +\input{tex/src/appendix-existing-tools.tex} +\input{tex/src/appendix-existing-solutions.tex} diff --git a/tex/src/appendix-existing-solutions.tex b/tex/src/appendix-existing-solutions.tex new file mode 100644 index 0000000..d80c0b4 --- /dev/null +++ b/tex/src/appendix-existing-solutions.tex @@ -0,0 +1,443 @@ +\section{Survey of common existing reproducible workflows} +\label{appendix:existingsolutions} + +As reviewed in the introduction, the problem of reproducibility has received a lot of attention over the last three decades and various solutions have already been proposed. +The core principles that many of the existing solutions (including Maneage) aim to achieve are nicely summarized by the FAIR principles \citeappendix{wilkinson16}. +In this appendix, some of the solutions are reviewed. +The solutions are based on an evolving software landscape, therefore they are ordered by date: when the project has a webpage, the year of its first release is used for the sorting, otherwise their paper's publication year is used. + +For each solution, we summarize its methodology and discuss how it relates to the criteria proposed in this paper. +Freedom of the software/method is a core concept behind scientific reproducibility, as opposed to industrial reproducibility where a black box is acceptable/desirable. +Therefore proprietary solutions like Code Ocean\footnote{\inlinecode{\url{https://codeocean.com}}} or Nextjournal\footnote{\inlinecode{\url{https://nextjournal.com}}} will not be reviewed here. +Other studies have also attempted to review existing reproducible solutions, foro example \citeappendix{konkol20}. + +\subsection{Suggested rules, checklists, or criteria} +Before going into the various implementations, it is also useful to review existing suggested rules, checklists or criteria for computationally reproducible research. + +All the cases below are primarily targetted to immediate reproducibility and do not consider longevity explicitly. +Therefore, they lack a strong/clear completeness criterion (they mainly only suggest, rather than require, the recording of versions, and their ultimate suggestion of storing the full binary OS in a binary VM or container is problematic (as mentioned in \ref{appendix:independentenvironment} and \citeappendix{oliveira18}). + +Sandve et al. \citeappendix{sandve13} propose ``ten simple rules for reproducible computational research'' that can be applied in any project. +Generally, these are very similar to the criteria proposed here and follow a similar spirit, but they do not provide any actual research papers following up all those points, nor do they provide a proof of concept. +The Popper convention \citeappendix{jimenez17} also provides a set of principles that are indeed generally useful, among which some are common to the criteria here (for example, automatic validation, and, as in Maneage, the authors suggest providing a template for new users), +but the authors do not include completeness as a criterion nor pay attention to longevity (Popper itself is written in Python with many dependencies, and its core operating language has already changed once). +For more on Popper, please see Section \ref{appendix:popper}. + +For improved reproducibility in Jupyter notebook users, \citeappendix{rule19} propose ten rules to improve reproducibility and also provide links to example implementations. +These can be very useful for users of Jupyter, but are not generic for non-Jupyter-based computational projects. +Some criteria (which are indeed very good in a more general context) do not directly relate to reproducibility, for example their Rule 1: ``Tell a Story for an Audience''. +Generally, as reviewed in Sections \ref{sec:longevityofexisting} and \ref{appendix:jupyter}, Jupyter itself has many issues regarding reproducibility. + +To create Docker images, N\"ust et al. propose ``ten simple rules'' in \citeappendix{nust20}. +They recommend some issues that can indeed help increase the quality of Docker images and their production/usage, such as their rule 7 to ``mount datasets [only] at run time'' to separate the computational environment from the data. +However, long-term reproducibility of the images is not included as a criterion by these authors. +For example, they recommend using base operating systems, with version identification limited to a single brief identifier such as \inlinecode{ubuntu:18.04}, which has a serious problem with longevity issues (Section \ref{sec:longevityofexisting}). +Furthermore, in their proof-of-concept Dockerfile (listing 1), \inlinecode{rocker} is used with a tag (not a digest), which can be problematic due to the high risk of ambiguity (as discussed in Section \ref{appendix:containers}). + +\subsection{Reproducible Electronic Documents, RED (1992)} +\label{appendix:red} + +RED\footnote{\inlinecode{\url{http://sep.stanford.edu/doku.php?id=sep:research:reproducible}}} is the first attempt that we could find on doing reproducible research, see \cite{claerbout1992,schwab2000}. +It was developed within the Stanford Exploration Project (SEP) for Geophysics publications. +Their introductions on the importance of reproducibility, resonate a lot with today's environment in computational sciences. +In particular the heavy investment one has to make in order to re-do another scientist's work, even in the same team. +RED also influenced other early reproducible works, for example \citeappendix{buckheit1995}. + +To orchestrate the various figures/results of a project, from 1990, they used ``Cake'' \citeappendix{somogyi87}, a dialect of Make, for more on Make, see Appendix \ref{appendix:jobmanagement}. +As described in \cite{schwab2000}, in the latter half of that decade, they moved to GNU Make, which was much more commonly used, developed and came with a complete and up-to-date manual. +The basic idea behind RED's solution was to organize the analysis as independent steps, including the generation of plots, and organizing the steps through a Makefile. +This enabled all the results to be re-executed with a single command. +Several basic low-level Makefiles were included in the high-level/central Makefile. +The reader/user of a project had to manually edit the central Makefile and set the variable \inlinecode{RESDIR} (result dir), this is the directory where built files are kept. +Afterwards, the reader could set which figures/parts of the project to reproduce by manually adding its name in the central Makefile, and running Make. + +At the time, Make was already practiced by individual researchers and projects as a job orchestration tool, but SEP's innovation was to standardize it as an internal policy, and define conventions for the Makefiles to be consistent across projects. +This enabled new members to benefit from the already existing work of previous team members (who had graduated or moved to other jobs). +However, RED only used the existing software of the host system, it had no means to control them. +Therefore, with wider adoption, they confronted a ``versioning problem'' where the host's analysis software had different versions on different hosts, creating different results, or crashing \citeappendix{fomel09}. +Hence in 2006 SEP moved to a new Python-based framework called Madagascar, see Appendix \ref{appendix:madagascar}. + + + + + +\subsection{Apache Taverna (2003)} +\label{appendix:taverna} +Apache Taverna\footnote{\inlinecode{\url{https://taverna.incubator.apache.org}}} \citeappendix{oinn04} is a workflow management system written in Java with a graphical user interface which is still being developed. +A workflow is defined as a directed graph, where nodes are called ``processors''. +Each Processor transforms a set of inputs into a set of outputs and they are defined in the Scufl language (an XML-based language, were each step is an atomic task). +Other components of the workflow are ``Data links'' and ``Coordination constraints''. +The main user interface is graphical, where users move processors in the given space and define links between their inputs outputs (manually constructing a lineage like Figure \ref{fig:datalineage}). +Taverna is only a workflow manager and is not integrated with a package manager, hence the versions of the used software can be different in different runs. +\citeappendix{zhao12} have studied the problem of workflow decays in Taverna. + + + + + +\subsection{Madagascar (2003)} +\label{appendix:madagascar} +Madagascar\footnote{\inlinecode{\url{http://ahay.org}}} \citeappendix{fomel13} is a set of extensions to the SCons job management tool (reviewed in \ref{appendix:scons}). +Madagascar is a continuation of the Reproducible Electronic Documents (RED) project that was discussed in Appendix \ref{appendix:red}. +Madagascar has been used in the production of hundreds of research papers or book chapters\footnote{\inlinecode{\url{http://www.ahay.org/wiki/Reproducible_Documents}}}, 120 prior to \citeappendix{fomel13}. + +Madagascar does include project management tools in the form of SCons extensions. +However, it is not just a reproducible project management tool. +It is primarily a collection of analysis programs and tools to interact with RSF files, and plotting facilities. +The Regularly Sampled File (RSF) file format\footnote{\inlinecode{\url{http://www.ahay.org/wiki/Guide\_to\_RSF\_file\_format}}} is a custom plain-text file that points to the location of the actual data files on the filesystem and acts as the intermediary between Madagascar's analysis programs. +For example in our test of Madagascar 3.0.1, it installed 855 Madagascar-specific analysis programs (\inlinecode{PREFIX/bin/sf*}). +The analysis programs mostly target geophysical data analysis, including various project specific tools: more than half of the total built tools are under the \inlinecode{build/user} directory which includes names of Madagascar users. + +Besides the location or contents of the data, RSF also contains name/value pairs that can be used as options to Madagascar programs, which are built with inputs and outputs of this format. +Since RSF contains program options also, the inputs and outputs of Madagascar's analysis programs are read from, and written to, standard input and standard output. + +In terms of completeness, as long as the user only uses Madagascar's own analysis programs, it is fairly complete at a high level (not lower-level OS libraries). +However, this comes at the expense of a large amount of bloatware (programs that one project may never need, but is forced to build). +Also, the linking between the analysis programs (of a certain user at a certain time) and future versions of that program (that is updated in time) is not immediately obvious. +Madagascar could have been more useful to a larger community if the workflow components were maintained as a separate project compared to the analysis components. + +\subsection{GenePattern (2004)} +\label{appendix:genepattern} +GenePattern\footnote{\inlinecode{\url{https://www.genepattern.org}}} \citeappendix{reich06} (first released in 2004) is a client-server software containing many common analysis functions/modules, primarily focused for Gene studies. +Although its highly focused to a special research field, it is reviewed here because its concepts/methods are generic, and in the context of this paper. + +Its server-side software is installed with fixed software packages that are wrapped into GenePattern modules. +The modules are used through a web interface, the modern implementation is GenePattern Notebook \citeappendix{reich17}. +It is an extension of the Jupyter notebook (see Appendix \ref{appendix:editors}), which also has a special ``GenePattern'' cell that will connect to GenePattern servers for doing the analysis. +However, the wrapper modules just call an existing tool on the host system. +Given that each server may have its own set of installed software, the analysis may differ (or crash) when run on different GenePattern servers, hampering reproducibility. + +%% GenePattern shutdown announcement (although as of November 2020, it does not open any more!): https://www.genepattern.org/blog/2019/10/01/the-genomespace-project-is-ending-on-november-15-2019 +The primary GenePattern server was active since 2008 and had 40,000 registered users with 2000 to 5000 jobs running every week \citeappendix{reich17}. +However, it was shut down on November 15th 2019 due to end of funding. +All processing with this sever has stopped, and any archived data on it has been deleted. +Since GenePattern is free software, there are alternative public servers to use, so hopefully work on it will continue. +However, funding is limited and those servers may face similar funding problems. +This is a very nice example of the fragility of solutions that depend on archiving and running the research codes with high-level research products (including data, binary/compiled code which are expensive to keep) in one place. + + + + + +\subsection{Kepler (2005)} +Kepler\footnote{\inlinecode{\url{https://kepler-project.org}}} \citeappendix{ludascher05} is a Java-based Graphic User Interface workflow management tool. +Users drag-and-drop analysis components, called ``actors'', into a visual, directional graph, which is the workflow (similar to Figure \ref{fig:datalineage}). +Each actor is connected to others through the Ptolemy II\footnote{\inlinecode{\url{https://ptolemy.berkeley.edu}}} \citeappendix{eker03}. +In many aspects, the usage of Kepler and its issues for long-term reproducibility is like Apache Taverna (see Section \ref{appendix:taverna}). + + + + + +\subsection{VisTrails (2005)} +\label{appendix:vistrails} + +VisTrails\footnote{\inlinecode{\url{https://www.vistrails.org}}} \citeappendix{bavoil05} was a graphical workflow managing system. +According to its webpage, VisTrails maintainance has stopped since May 2016, its last Git commit, as of this writing, was in November 2017. +However, given that it was well maintained for over 10 years is an achievement. + +VisTrails (or ``visualization trails'') was initially designed for managing visualizations, but later grew into a generic workflow system with meta-data and provenance features. +Each analysis step, or module, is recorded in an XML schema, which defines the operations and their dependencies. +The XML attributes of each module can be used in any XML query language to find certain steps (for example those that used a certain command). +Since the main goal was visualization (as images), apparently its primary output is in the form of image spreadsheets. +Its design is based on a change-based provenance model using a custom VisTrails provenance query language (vtPQL), for more see \citeappendix{scheidegger08}. +Since XML is a plane text format, as the user inspects the data and makes changes to the analysis, the changes are recorded as ``trails'' in the project's VisTrails repository that operates very much like common version control systems (see Appendix \ref{appendix:versioncontrol}). +. +However, even though XML is in plain text, it is very hard to edit manually. +VisTrails therefore provides a graphic user interface with a visual representation of the project's inter-dependent steps (similar to Figure \ref{fig:datalineage}). +Besides the fact that it is no longer maintained, VisTrails does not control the software that is run, it only controls the sequence of steps that they are run in. + + + + + +\subsection{Galaxy (2010)} +\label{appendix:galaxy} + +Galaxy\footnote{\inlinecode{\url{https://galaxyproject.org}}} is a web-based Genomics workbench \citeappendix{goecks10}. +The main user interface are ``Galaxy Pages'', which does not require any programming: users simply use abstract ``tools'' which are a wrappers over command-line programs. +Therefore the actual running version of the program can be hard to control across different Galaxy servers. +Besides the automatically generated metadata of a project (which include version control, or its history), users can also tag/annotate each analysis step, describing its intent/purpose. +Besides some small differences Galaxy seems very similar to GenePattern (Appendix \ref{appendix:genepattern}), so most of the same points there apply here too (including the very large cost of maintining such a system). + + + + + +\subsection{Image Processing On Line journal, IPOL (2010)} +\label{appendix:ipol} +The IPOL journal\footnote{\inlinecode{\url{https://www.ipol.im}}} \citeappendix{limare11} (first published article in July 2010) publishes papers on image processing algorithms as well as the the full code of the proposed algorithm. +An IPOL paper is a traditional research paper, but with a focus on implementation. +The published narrative description of the algorithm must be detailed to a level that any specialist can implement it in their own programming language (extremely detailed). +The author's own implementation of the algorithm is also published with the paper (in C, C++ or MATLAB), the code must be commented well enough and link each part of it with the relevant part of the paper. +The authors must also submit several example datasets/scenarios. +The referee is expected to inspect the code and narrative, confirming that they match with each other, and with the stated conclusions of the published paper. +After publication, each paper also has a ``demo'' button on its webpage, allowing readers to try the algorithm on a web-interface and even provide their own input. + +IPOL has grown steadily over the last 10 years, publishing 23 research articles in 2019 alone. +We encourage the reader to visit its webpage and see some of its recent papers and their demos. +The reason it can be so thorough and complete is its very narrow scope (low-level image processing algorithms), where the published algorithms are highly atomic, not needing significant dependencies (beyond input/output of well-known formats), allowing the referees and readers to go deeply into each implemented algorithm. +In fact, high-level languages like Perl, Python or Java are not acceptable in IPOL precisely because of the additional complexities, such as dependencies, that they require. +However, many data-intensive projects commonly involve dozens of high-level dependencies, with large and complex data formats and analysis, so this solution is not scalable. + +IPOL thus fails on our Scalability criteria. +Furthermore, by not publishing/archiving each paper's version control history or directly linking the analysis and produced paper, it fails Criterias 6 and 7. +Note that on the webpage, it is possible to change parameters, but that will not affect the produced PDF. +A paper written in Maneage (the proof-of-concept solution presented in this paper) could be scrutinised at a similar detailed level to IPOL, but for much more complex research scenarios, involving hundreds of dependencies and complex processing of the data. + + + + + + +\subsection{WINGS (2010)} +\label{appendix:wings} + +WINGS\footnote{\inlinecode{\url{https://wings-workflows.org}}} \citeappendix{gil10} is an automatic workflow generation algorithm. +It runs on a centralized web server, requiring many dependencies (such that it is recommended to download Docker images). +It allows users to define various workflow components (for example datasets, analysis components and etc), with high-level goals. +It then uses selection and rejection algorithms to find the best components using a pool of analysis components that can satisfy the requested high-level constraints. +\tonote{Read more about this} + + + + + +\subsection{Active Papers (2011)} +\label{appendix:activepapers} +Active Papers\footnote{\inlinecode{\url{http://www.activepapers.org}}} attempts to package the code and data of a project into one file (in HDF5 format). +It was initially written in Java because its compiled byte-code outputs in JVM are portable on any machine \citeappendix{hinsen11}. +However, Java is not a commonly used platform today, hence it was later implemented in Python \citeappendix{hinsen15}. + +In the Python version, all processing steps and input data (or references to them) are stored in a HDF5 file. +However, it can only account for pure-Python packages using the host operating system's Python modules \tonote{confirm this!}. +When the Python module contains a component written in other languages (mostly C or C++), it needs to be an external dependency to the Active Paper. + +As mentioned in \citeappendix{hinsen15}, the fact that it relies on HDF5 is a caveat of Active Papers, because many tools are necessary to merely open it. +Downloading the pre-built HDF View binaries (provided by the HDF group) is not possible anonymously/automatically (login is required). +Installing it using the Debian or Arch Linux package managers also failed due to dependencies in our trials. +Furthermore, as a high-level data format HDF5 evolves very fast, for example HDF5 1.12.0 (February 29th, 2020) is not usable with older libraries provided by the HDF5 team. % maybe replace with: February 29\textsuperscript{th}, 2020? + +While data and code are indeed fundamentally similar concepts technically\citeappendix{hinsen16}, they are used by humans differently due to their volume: the code of a large project involving Terabytes of data can be less than a megabyte. +Hence, storing code and data together becomes a burden when large datasets are used, this was also acknowledged in \citeappendix{hinsen15}. +Also, if the data are proprietary (for example medical patient data), the data must not be released, but the methods that were applied to them can be published. +Furthermore, since all reading and writing is done in the HDF5 file, it can easily bloat the file to very large sizes due to temporary/reproducible files, and its necessary to remove/dummify them, thus complicating the code, making it hard to read. +For example the Active Papers HDF5 file of \citeappendix[in \href{https://doi.org/10.5281/zenodo.2549987}{zenodo.2549987}]{kneller19} is 1.8 giga-bytes. + +In many scenarios, peers just want to inspect the processing by reading the code and checking a very special part of it (one or two lines, just to see the option values to one step for example). +They do not necessarily need to run it, or obtaining the datasets. +Hence the extra volume for data, and obscure HDF5 format that needs special tools for reading plain text code is a major hinderance. + + + + + +\subsection{Collage Authoring Environment (2011)} +\label{appendix:collage} +The Collage Authoring Environment \citeappendix{nowakowski11} was the winner of Elsevier Executable Paper Grand Challenge \citeappendix{gabriel11}. +It is based on the GridSpace2\footnote{\inlinecode{\url{http://dice.cyfronet.pl}}} distributed computing environment\tonote{find citation}, which has a web-based graphic user interface. +Through its web-based interface, viewers of a paper can actively experiment with the parameters of a published paper's displayed outputs (for example figures). +\tonote{See how it containerizes the software environment} + + + + + +\subsection{SHARE (2011)} +\label{appendix:SHARE} +SHARE\footnote{\inlinecode{\url{https://is.ieis.tue.nl/staff/pvgorp/share}}} \citeappendix{vangorp11} is a web portal that hosts virtual machines (VMs) for storing the environment of a research project. +The top project webpage above is still active, however, the virtual machines and SHARE system have been removed since 2019, probably due to the large volume and high maintainence cost of the VMs. + +SHARE was recognized as second position in the Elsevier Executable Paper Grand Challenge \citeappendix{gabriel11}. +Simply put, SHARE is just a VM that users can download and run. +The limitations of VMs for reproducibility were discussed in Appendix \ref{appendix:virtualmachines}, and the SHARE system does not specify any requirements on making the VM itself reproducible. + + + + + +\subsection{Verifiable Computational Result, VCR (2011)} +\label{appendix:verifiableidentifier} +A ``verifiable computational result''\footnote{\inlinecode{\url{http://vcr.stanford.edu}}} is an output (table, figure, or etc) that is associated with a ``verifiable result identifier'' (VRI), see \citeappendix{gavish11}. +It was awarded the third prize in the Elsevier Executable Paper Grand Challenge \citeappendix{gabriel11}. + +A VRI is created using tags within the programming source that produced that output, also recording its version control or history. +This enables exact identification and citation of results. +The VRIs are automatically generated web-URLs that link to public VCR repositories containing the data, inputs and scripts, that may be re-executed. +According to \citeappendix{gavish11}, the VRI generation routine has been implemented in MATLAB, R and Python, although only the MATLAB version was available during the writing of this paper. +VCR also has special \LaTeX{} macros for loading the respective VRI into the generated PDF. + +Unfortunately most parts of the webpage are not complete at the time of this writing. +The VCR webpage contains an example PDF\footnote{\inlinecode{\url{http://vcr.stanford.edu/paper.pdf}}} that is generated with this system, however, the linked VCR repository\footnote{\inlinecode{\url{http://vcr-stat.stanford.edu}}} does not exist at the time of this writing. +Finally, the date of the files in the MATLAB extension tarball are set to 2011, hinting that probably VCR has been abandoned soon after the publication of \citeappendix{gavish11}. + + + + + +\subsection{SOLE (2012)} +\label{appendix:sole} +SOLE (Science Object Linking and Embedding) defines ``science objects'' (SOs) that can be manually linked with phrases of the published paper \citeappendix{pham12,malik13}. +An SO is any code/content that is wrapped in begin/end tags with an associated type and name. +For example special commented lines in a Python, R or C program. +The SOLE command-line program parses the tagged file, generating metadata elements unique to the SO (including its URI). +SOLE also supports workflows as Galaxy tools \citeappendix{goecks10}. + +For reproducibility, \citeappendix{pham12} suggest building a SOLE-based project in a virtual machine, using any custom package manager that is hosted on a private server to obtain a usable URI. +However, as described in Appendices \ref{appendix:independentenvironment} and \ref{appendix:packagemanagement}, unless virtual machines are built with robust package managers, this is not a sustainable solution (the virtual machine itself is not reproducible). +Also, hosting a large virtual machine server with fixed IP on a hosting service like Amazon (as suggested there) for every project in perpetuity will be very expensive. +The manual/artificial definition of tags to connect parts of the paper with the analysis scripts is also a caveat due to human error and incompleteness (tags the authors may not consider important, but may be useful later). +In Maneage, instead of artificial/commented tags directly link the analysis input and outputs to the paper's text automatically. + + + +\subsection{Sumatra (2012)} +Sumatra\footnote{\inlinecode{\url{http://neuralensemble.org/sumatra}}} \citeappendix{davison12} attempts to capture the environment information of a running project. +It is written in Python and is a command-line wrapper over the analysis script. +By controlling a project at running-time, Sumatra is able to capture the environment it was run in. +The captured environment can be viewed in plain text or a web interface. +Sumatra also provides \LaTeX/Sphinx features, which will link the paper with the project's Sumatra database. +This enables researchers to use a fixed version of a project's figures in the paper, even at later times (while the project is being developed). + +The actual code that Sumatra wraps around, must itself be under version control, and it does not run if there is non-committed changes (although its not clear what happens if a commit is amended). +Since information on the environment has been captured, Sumatra is able to identify if it has changed since a previous run of the project. +Therefore Sumatra makes no attempt at storing the environment of the analysis as in Sciunit (see Appendix \ref{appendix:sciunit}), but its information. +Sumatra thus needs to know the language of the running program and is not generic. +It just captures the environment, it does not store \emph{how} that environment was built. + + + + + +\subsection{Research Object (2013)} +\label{appendix:researchobject} + +The Research object\footnote{\inlinecode{\url{http://www.researchobject.org}}} is collection of meta-data ontologies, to describe aggregation of resources, or workflows, see \citeappendix{bechhofer13} and \citeappendix{belhajjame15}. +It thus provides resources to link various workflow/analysis components (see Appendix \ref{appendix:existingtools}) into a final workflow. + +\citeappendix{bechhofer13} describes how a workflow in Apache Taverna (Appendix \ref{appendix:taverna}) can be translated into research objects. +The important thing is that the research object concept is not specific to any special workflow, it is just a metadata bundle/standard which is only as robust in reproducing the result as the running workflow. + + + + +\subsection{Sciunit (2015)} +\label{appendix:sciunit} +Sciunit\footnote{\inlinecode{\url{https://sciunit.run}}} \citeappendix{meng15} defines ``sciunit''s that keep the executed commands for an analysis and all the necessary programs and libraries that are used in those commands. +It automatically parses all the executables in the script, and copies them, and their dependency libraries (down to the C library), into the sciunit. +Because the sciunit contains all the programs and necessary libraries, its possible to run it readily on other systems that have a similar CPU architecture. +Sciunit was originally written in Python 2 (which reached its end-of-life in January 1st, 2020). +Therefore Sciunit2 is a new implementation in Python 3. + +The main issue with Sciunit's approach is that the copied binaries are just black boxes: it is not possible to see how the used binaries from the initial system were built. +This is a major problem for scientific projects: in principle (not knowing how they programs were built) and in practice (archiving a large volume sciunit for every step of an analysis requires a lot of storage space). + + + + + +\subsection{Umbrella (2015)} +Umbrella \citeappendix{meng15b} is a high-level wrapper script for isolating the environment of an analysis. +The user specifies the necessary operating system, and necessary packages and analysis steps in variuos JSON files. +Umbrella will then study the host operating system and the various necessary inputs (including data and software) through a process similar to Sciunits mentioned above to find the best environment isolator (maybe using Linux containerization, containers or VMs). +We could not find a URL to the source software of Umbrella (no source code repository is mentioned in the papers we reviewed above), but from the descriptions in \citeappendix{meng17}, it is written in Python 2.6 (which is now \new{deprecated}). + + + + + +\subsection{ReproZip (2016)} +ReproZip\footnote{\inlinecode{\url{https://www.reprozip.org}}} \citeappendix{chirigati16} is a Python package that is designed to automatically track all the necessary data files, libraries and environment variables into a single bundle. +The tracking is done at the kernel system-call level, so any file that is accessed during the running of the project is identified. +The tracked files can be packaged into a \inlinecode{.rpz} bundle that can then be unpacked into another system. + +ReproZip is therefore very good to take a ``snapshot'' of the running environment into a single file. +The bundle can become very large if large, or many datasets, are used or if the software evironment is complex (many dependencies). +Since it copies the binary software libraries, it can only be run on systems with a similar CPU architecture to the original. +Furthermore, ReproZip just copies the binary/compiled files used in a project, it has no way to know how those software were built. +As mentioned in this paper, and also \citeappendix{oliveira18} the question of ``how'' the environment was built is critical for understanding the results and simply having the binaries cannot necessarily be useful. + +For the data, it is similarly not possible to extract which data server they came from. +Hence two projects that each use a 1-terabyte dataset will need a full copy of that same 1-terabyte file in their bundle, making long term preservation extremely expensive. + + + + + +\subsection{Binder (2017)} +Binder\footnote{\inlinecode{\url{https://mybinder.org}}} is used to containerize already existing Jupyter based processing steps. +Users simply add a set of Binder-recognized configuration files to their repository and Binder will build a Docker image and install all the dependencies inside of it with Conda (the list of necessary packages comes from Conda). +One good feature of Binder is that the imported Docker image must be tagged (something like a checksum). +This will ensure that future/latest updates of the imported Docker image are not mistakenly used. +However, it does not make sure that the dockerfile used by the imported Docker image follows a similar convention also. +Binder is used by \citeappendix{jones19}. + + + + + +\subsection{Gigantum (2017)} +%% I took the date from their PiPy page, where the first version 0.1 was published in November 2016. +Gigantum\footnote{\inlinecode{\url{https://gigantum.com}}} is a client/server system, in which the client is a web-based (graphical) interface that is installed as ``Gigantum Desktop'' within a Docker image. +Gigantum uses Docker containers for an independent environment, Conda (or Pip) to install packages, Jupyter notebooks to edit and run code, and Git to store its history. +Simply put, its a high-level wrapper for combining these components. +Internally, a Gigantum project is organized as files in a directory that can be opened without their own client. +The file structure (which is under version control) includes codes, input data and output data. +As acknowledged on their own webpage, this greatly reduces the speed of Git operations, transmitting, or archiving the project. +Therefore there are size limits on the dataset/code sizes. +However, there is one directory which can be used to store files that must not be tracked. + + + + +\subsection{Popper (2017)} +\label{appendix:popper} +Popper\footnote{\inlinecode{\url{https://falsifiable.us}}} is a software implementation of the Popper Convention \citeappendix{jimenez17}. +The Popper team's own solution is through a command-line program called \inlinecode{popper}. +The \inlinecode{popper} program itself is written in Python. +However, job management was initially based on the HashiCorp configuration language (HCL) because HCL was used by ``GitHub Actions'' to manage workflows. +Moreover, from October 2019 Github changed to a custom YAML-based languguage, so Popper also deprecated HCL. +This is an important issue when low-level choices are based on service providers. + +To start a project, the \inlinecode{popper} command-line program builds a template, or ``scaffold'', which is a minimal set of files that can be run. +However, as of this writing, the scaffold is not complete: it lacks a manuscript and validation of outputs (as mentioned in the convention). +By default Popper runs in a Docker image (so root permissions are necessary and reproducible issues with Docker images have been discussed above), but Singularity is also supported. +See Appendix \ref{appendix:independentenvironment} for more on containers, and Appendix \ref{appendix:highlevelinworkflow} for using high-level languages in the workflow. + +Popper does not comply with the completeness, minimal complexity and including-narrative criteria. +Moreover, the scaffold that is provided by Popper is an output of the program that is not directly under version control. +Hence, tracking future changes in Popper and how they relate to the high-level projects that depend on it will be very hard. +In Maneage, the same \inlinecode{maneage} git branch is shared by the developers and users; any new feature or change in Maneage can thus be directly tracked with Git when the high-level project merges their branch with Maneage. + + +\subsection{Whole Tale (2017)} +\label{appendix:wholetale} + +Whole Tale\footnote{\inlinecode{\url{https://wholetale.org}}} is a web-based platform for managing a project and organizing data provenance, see \citeappendix{brinckman17} +It uses online editors like Jupyter or RStudio (see Appendix \ref{appendix:editors}) that are encapsulated in a Docker container (see Appendix \ref{appendix:independentenvironment}). + +The web-based nature of Whole Tale's approach, and its dependency on many tools (which have many dependencies themselves) is a major limitation for future reproducibility. +For example, when following their own tutorial on ``Creating a new tale'', the provided Jupyter notebook could not be executed because of a dependency problem. +This was reported to the authors as issue 113\footnote{\inlinecode{\url{https://github.com/whole-tale/wt-design-docs/issues/113}}}, but as all the second-order dependencies evolve, its not hard to envisage such dependency incompatibilities being the primary issue for older projects on Whole Tale. +Furthermore, the fact that a Tale is stored as a binary Docker container causes two important problems: +1) it requires a very large storage capacity for every project that is hosted there, making it very expensive to scale if demand expands. +2) It is not possible to see how the environment was built accurately (when the Dockerfile uses \inlinecode{apt}). +This issue with Whole Tale (and generally all other solutions that only rely on preserving a container/VM) was also mentioned in \citeappendix{oliveira18}, for more on this, please see Appendix \ref{appendix:packagemanagement}. + + + + + +\subsection{Occam (2018)} +Occam\footnote{\inlinecode{\url{https://occam.cs.pitt.edu}}} \citeappendix{oliveira18} is web-based application to preserve software and its execution. +To achieve long-term reproducibility, Occam includes its own package manager (instructions to build software and their dependencies) to be in full control of the software build instructions, similar to Maneage. +Besides Nix or Guix (which are primarily a package manager that can also do job management), Occum has been the only solution in our survey here that attempts to be complete in this aspect. + +However it is incomplete from the perspective of requirements: it works within a Docker image (that requires root permissions) and currently only runs on Debian-based, Redhat-based and Arch-based GNU/Linux operating systems that respectively use the \inlinecode{apt}, \inlinecode{pacman} or \inlinecode{yum} package managers. +It is also itself written in Python (version 3.4 or above), hence it is not clear + +Furthermore, it also violates our complexity criteria because the instructions to build the software, their versions and etc are not immediately viewable or modifable by the user. +Occam contains its own JSON database for this that should be parsed with its own custom program. +The analysis phase of Occum is also through a drag-and-drop interface (similar to Taverna, Appendix \ref{appendix:taverna}) that is a web-based graphic user interface. +All the connections between various phases of an analysis need to be pre-defined in a JSON file and manually linked in the GUI. +Hence for complex data analysis operations with involve thousands of steps, it is not scalable. diff --git a/tex/src/appendix-existing-tools.tex b/tex/src/appendix-existing-tools.tex new file mode 100644 index 0000000..8ad1cc3 --- /dev/null +++ b/tex/src/appendix-existing-tools.tex @@ -0,0 +1,537 @@ +\section{Survey of existing tools for various phases} +\label{appendix:existingtools} + +Data analysis workflows (including those that aim for reproducibility) are commonly high-level frameworks which employ various lower-level components. +To help in reviewing existing reproducible workflow solutions in light of the proposed criteria in Appendix \ref{appendix:existingsolutions}, we first need to survey the most commonly employed lower-level tools. + +\subsection{Independent environment} +\label{appendix:independentenvironment} + +The lowest-level challenge of any reproducible solution is to avoid the differences between various run-time environments, to a desirable/certain level. +For example different hardware, operating systems, versions of existing dependencies, and etc. +Therefore any reasonable attempt at providing a reproducible workflow starts with isolating its running envionment from the host environment. +There are three general technologies that are used for this purpose and reviewed below: +1) Virtual machines, +2) Containers, +3) Independent build in host's file system. + +\subsubsection{Virtual machines} +\label{appendix:virtualmachines} +Virtual machines (VMs) host a binary copy of a full operating system that can be run on other operating systems. +This includes the lowest-level operating system component or the kernel. +VMs thus provide the ultimate control one can have over the run-time environment of an analysis. +However, the VM's kernel does not talk directly to the running hardware that is doing the analysis, it talks to a simulated hardware layer that is provided by the host's kernel. +Therefore, a process that is run inside a virtual machine can be much slower than one that is run on a native kernel. +An advantages of VMs is that they are a single file which can be copied from one computer to another, keeping the full environment within them if the format is recognized. +VMs are used by cloud service providers, enabling fully independent operating systems on their large servers (where the customer can have root access). + +VMs were used in solutions like SHARE \citeappendix{vangorp11} (which was awarded second prize in the Elsevier Executable Paper Grand Challenge of 2011 \citeappendix{gabriel11}), or in suggested reproducible papers like \citeappendix{dolfi14}. +However, due to their very large size, these are expensive to maintain, thus leading SHARE to discontinue its services in 2019. +The URL to the VM file \texttt{provenance\_machine.ova} that is mentioned in \citeappendix{dolfi14} is not currently accessible (we suspect that this is due to size and archival costs). + +\subsubsection{Containers} +\label{appendix:containers} +Containers also host a binary copy of a running environment, but do not have their own kernel. +Through a thin layer of low-level system libraries, programs running within a container talk directly with the host operating system kernel. +Otherwise, containers have their own independent software for everything else. +Therefore, they have much less overhead in hardware/CPU access. +Like VMs, users often choose an operating system for the container's independent operating system (most commonly GNU/Linux distributions which are free software). + +Below we review some of the most common container solutions: Docker and Singularity. + +\begin{itemize} +\item {\bf\small Docker containers:} Docker is one of the most popular tools today for keeping an independent analysis environment. + It is primarily driven by the need of software developers for reproducing a previous environment, where they have root access mostly on the ``cloud'' (which is just a remote VM). + A Docker container is composed of independent Docker ``images'' that are built with a \inlinecode{Dockerfile}. + It is possible to precisely version/tag the images that are imported (to avoid downloading the latest/different version in a future build). + To have a reproducible Docker image, it must be ensured that all the imported Docker images check their dependency tags down to the initial image which contains the C library. + + An important drawback of Docker for high performance scientific needs is that it runs as a daemon (a program that is always running in the background) with root permissions. + This is a major security flaw that discourages many high performance computing (HPC) facilities from providing it. + +\item {\bf\small Singularity:} Singularity \citeappendix{kurtzer17} is a single-image container (unlike Docker which is composed of modular/independent images). + Although it needs root permissions to be installed on the system (once), it does not require root permissions every time it is run. + Its main program is also not a daemon, but a normal program that can be stopped. + These features make it much safer for HPC administrators to install compared to Docker. + However, the fact that it requires root access for the initial install is still a hindrance for a typical project: if Singularity is not already present on the HPC, the user's science project cannot be run by a non-root user. + +\item {\bf\small Podman:} Podman uses the Linux kernel containerization features to enable containers without a daemon, and without root permissions. + It has a command-line interface very similar to Docker, but only works on GNU/Linux operating systems. +\end{itemize} + +Generally, VMs or containers are good solutions to reproducibly run/repeating an analysis in the short term (a couple of years). +However, their focus is to store the already-built (binary, non-human readable) software environment. +Because of this they will be large (many Gigabytes) and expensive to archive, download or access. +Recall the two examples above for VMs in Section \ref{appendix:virtualmachines}. But this is also valid for Docker images, as is clear from Dockerhub's recent decision to delete images of free accounts that have not been used for more than 6 months. +Meng \& Thain \citeappendix{meng17} also give similar reasons on why Docker images were not suitable in their trials. + +On a more fundamental level, VMs or contains do not store \emph{how} the core environment was built. +This information is usually in a third-party repository, and not necessarily inside container or VM file, making it hard (if not impossible) to track for future users. +This is a major problem when considering reproducibility which is also highlighted as a major issue in terms of long term reproducibility in \citeappendix{oliveira18}. + +The example of \cite{mesnard20} was previously mentioned in Section \ref{criteria}. +Another useful example is the \href{https://github.com/benmarwick/1989-excavation-report-Madjedbebe/blob/master/Dockerfile}{\inlinecode{Dockerfile}} of \citeappendix{clarkso15} (published in June 2015) which starts with \inlinecode{FROM rocker/verse:3.3.2}. +When we tried to build it (November 2020), the core downloaded image (\inlinecode{rocker/verse:3.3.2}, with image ``digest'' \inlinecode{sha256:c136fb0dbab...}) was created in October 2018 (long after the publication of that paper). +In principle, it is possible to investigate the difference between this new image and the old one that the authors used, but that would require a lot of effort and may not be possible where the changes are not available in a third public repository or not under version control. +In Docker, it is possible to retrieve the precise Docker image with its digest for example \inlinecode{FROM ubuntu:16.04@sha256:XXXXXXX} (where \inlinecode{XXXXXXX} is the digest, uniquely identifying the core image to be used), but we have not seen this often done in existing examples of ``reproducible'' \inlinecode{Dockerfiles}. + +The ``digest'' is specific to Docker repositories. +A more generic/longterm approach to ensure identical core OS components at a later time is to construct the containers or VMs with fixed/archived versions of the operating system ISO files. +ISO files are pre-built binary files with volumes of hundreds of megabytes and not containing their build instructions). +For example the archives of Debian\footnote{\inlinecode{\url{https://cdimage.debian.org/mirror/cdimage/archive/}}} or Ubuntu\footnote{\inlinecode{\url{http://old-releases.ubuntu.com/releases}}} provide older ISO files. + +The concept of containers (and the independent images that build them) can also be extended beyond just the software environment. +For example \citeappendix{lofstead19} propose a ``data pallet'' concept to containerize access to data and thus allow tracing data back wards to the application that produced them. + +In summary, containers or VMs are just a built product themselves. +If they are built properly (for example building a Maneage'd project inside a Docker container), they can be useful for immediate usage and fast moving of the project from one system to another. +With robust building, the container or VM can also be exactly reproduced later. +However, attempting to archive the actual binary container or VM files as a black box (not knowing the precise versions of the software in them, and \emph{how} they were built) is expensive, and will not be able to answer the most fundamental questions. + +\subsubsection{Independent build in host's file system} +\label{appendix:independentbuild} +The virtual machine and container solutions mentioned above, have their own independent file system. +Another approach to having an isolated analysis environment is to use the same filesystem as the host, but installing the project's software in a non-standrard, project-specific directory that does not interfere with the host. +Because the environment in this approach can be built in any custom location on the host, this solution generally does not require root permissions or extra low-level layers like containers or VMs. +However, ``moving'' the built product of such solutions from one computer to another is not generally as trivial as containers or VMs. +Examples of such third-party package managers (that are detached from the host OS's package manager) include Nix, GNU Guix, Python's Virtualenv package and Conda, among others. +Because it is highly intertwined with the way software are built and installed, third party package managers are described in more detail as part of Section \ref{appendix:packagemanagement}. + +Maneage (the solution proposed in this paper) also follows a similar approach of building and installing its own software environment within the the host's file system but without depending on it beyond the kernel. +However, unlike the third party package maneager mentioned above, based on the Completeness criteria above Maneage's package management is not detached from the specific research/analysis project: the instructions to build the full isolated software environment is maintained with the high-level analysis steps of the project, and the narrative paper/report of the project. + + + + + +\subsection{Package management} +\label{appendix:packagemanagement} + +Package management is the process of automating the build and installation of a software environment. +A package manager thus contains the following information on each software package that can be run automatically: the URL of the software's tarball, the other software that it possibly depends on, and how to configure and build it. +Package managers can be tied to specific operating systems at a very low level (like \inlinecode{apt} in Debian-based OSs). +Alternatively, there are third-party package managers which ca be installed on many OSs. +Both are discussed in more detail below. + +Package managers are the second component in any workflow that relies on containers or VMs for an independent environment, and the starting point in others that use the host's file system (as discussed above in Section \ref{appendix:independentenvironment}). +In this section, some common package managers are reviewed, in particular those that are most used by the reviewed reproducibility solutions of Appendix \ref{appendix:existingsolutions}. +For a more comprehensive list of existing package managers, see \href{https://en.wikipedia.org/wiki/List_of_software_package_management_systems}{Wikipedia}. +Note that we are not including package managers that are specific to one language, for example \inlinecode{pip} (for Python) or \inlinecode{tlmgr} (for \LaTeX). + + + +\subsubsection{Operating system's package manager} +The most commonly used package managers are those of the host operating system, for example \inlinecode{apt} or \inlinecode{yum} respectively on Debian-based, or RedHat-based GNU/Linux operating systems, \inlinecode{pkg} in FreeBSD, among many others in other OSes. + +These package managers are tightly intertwined with the operating system: they also include the building and updating of the core kernel and the C library. +Because they are part of the OS, they also commonly require root permissions. +Also, it is usually only possible to have one version/configuration of a software at any moment and downgrading versions for one project, may conflict with other projects, or even cause problems in the OS. +Hence if two projects need different versions of a software, it is not possible to work on them at the same time in the OS. + +When a container or virtual machine (see Appendix \ref{appendix:independentenvironment}) is used for each project, it is common for projects to use the containerized operating system's package manager. +However, it is important to remember that operating system package managers are not static: software are updated on their servers. +Hence, simply running \inlinecode{apt install gcc}, will install different versions of the GNU Compiler Collection (GCC) based on the version of the OS and when it has been run. +Requesting a special version of that special software does not fully address the problem because the package managers also download and install its dependencies. +Hence a fixed version of the dependencies must also be specified. + +In robust package managers like Debian's \inlinecode{apt} it is possible to fully control (and later reproduce) the build environment of a high-level software. +Debian also archives all packaged high-level software in its Snapshot\footnote{\inlinecode{\url{https://snapshot.debian.org/}}} service since 2005 which can be used to build the higher-level software environment on an older OS \citeappendix{aissi20}. +Hence it is indeed theoretically possible to reproduce the software environment only using archived operating systems and their own package managers, but unfortunately we have not seen it practiced in scientific papers/projects. + +In summary, the host OS package managers are primarily meant for the operating system components or very low-level components. +Hence, many robust reproducible analysis solutions (reviewed in Appendix \ref{appendix:existingsolutions}) do not use the host's package manager, but an independent package manager, like the ones below discussed below. + +\subsubsection{Packaging with Linux containerization} +Once a software is packaged as an AppImage\footnote{\inlinecode{\url{https://appimage.org}}}, Flatpak\footnote{\inlinecode{\url{https://flatpak.org}}} or Snap\footnote{\inlinecode{\url{https://snapcraft.io}}} the software's binary product and all its dependencies (not including the core C library) are packaged into one file. +This makes it very easy to move that single software's built product to newer systems. +However, because the C library is not included, it can fail on older systems. +Moreover, these are designed for the Linux kernel (using its containerization features) and can thus only be run on GNU/Linux operating systems. + +\subsubsection{Nix or GNU Guix} +\label{appendix:nixguix} +Nix \citeappendix{dolstra04} and GNU Guix \citeappendix{courtes15} are independent package managers that can be installed and used on GNU/Linux operating systems, and macOS (only for Nix, prior to macOS Catalina). +Both also have a fully functioning operating system based on their packages: NixOS and ``Guix System''. +GNU Guix is based on the same principles of Nix but implemented differencely, so we focus the review here on Nix. + +The Nix approach to package management is unique in that it allows exact dependency tracking of all the dependencies, and allows for multiple versions of a software, for more details see \citeappendix{dolstra04}. +In summary, a unique hash is created from all the components that go into the building of the package. +That hash is then prefixed to the software's installation directory. +As an example from \citeappendix{dolstra04}: if a certain build of GNU C Library 2.3.2 has a hash of \inlinecode{8d013ea878d0}, then it is installed under \inlinecode{/nix/store/8d013ea878d0-glibc-2.3.2} and all software that are compiled with it (and thus need it to run) will link to this unique address. +This allows for multiple versions of the software to co-exist on the system, while keeping an accurate dependency tree. + +As mentioned in \citeappendix{courtes15}, one major caveat with using these package managers is that they require a daemon with root privileges. +This is necessary ``to use the Linux kernel container facilities that allow it to isolate build processes and maximize build reproducibility''. +This is because the focus in Nix or Guix is to create bit-wise reproducible software binaries and this is necessary in the security or development perspectives. +However, in a non-computer-science analysis (for example natural sciences), the main aim is reproducibile \emph{results} that can also be created with the same software version that may not be bitwise identical (for example when they are installed in other locations, because the installation location is hardcoded in the software binary). + +Finally, while Guix and Nix do allow preciesly reproducible environments, it requires extra effort. +For example simply running \inlinecode{guix install gcc} will install the most recent version of GCC that can be different at different times. +Hence, similar to the discussion in host operating system package managers, it is up to the user to ensure that their created environment is recorded properly for reproducibility in the future. +Generally, this is a major limitation of projects that rely on detached package managers for building their software, including the other tools mentioned below. + +\subsubsection{Conda/Anaconda} +\label{appendix:conda} +Conda is an independent package manager that can be used on GNU/Linux, macOS, or Windows operating systems, although all software packages are not available in all operating systems. +Conda is able to maintain an approximately independent environment on an operating system without requiring root access. + +Conda tracks the dependencies of a package/environment through a YAML formatted file, where the necessary software and their acceptable versions are listed. +However, it is not possible to fix the versions of the dependencies through the YAML files alone. +This is thoroughly discussed under issue 787 (in May 2019) of \inlinecode{conda-forge}\footnote{\url{https://github.com/conda-forge/conda-forge.github.io/issues/787}}. +In that discussion, the authors of \citeappendix{uhse19} report that the half-life of their environment (defined in a YAML file) is 3 months, and that at least one of their their dependencies breaks shortly after this period. +The main reply they got in the discussion is to build the Conda environment in a container, which is also the suggested solution by \citeappendix{gruning18}. +However, as described in Appendix \ref{appendix:independentenvironment} containers just hide the reproducibility problem, they do not fix it: containers are not static and need to evolve (i.e., re-built) with the project. +Given these limitations, \citeappendix{uhse19} are forced to host their conda-packaged software as tarballs on a separate repository. + +Conda installs with a shell script that contains a binary-blob (+500 megabytes, embedded in the shell script). +This is the first major issue with Conda: from the shell script, it is not clear what is in this binary blob and what it does. +After installing Conda in any location, users can easily activate that environment by loading a special shell script into their shell. +However, the resulting environment is not fully independent of the host operating system as described below: + +\begin{itemize} +\item The Conda installation directory is present at the start of environment variables like \inlinecode{PATH} (which is used to find programs to run) and other such environment variables. + However, the host operating system's directories are also appended afterwards. + Therefore, a user, or script may not notice that a software that is being used is actually coming from the operating system, not the controlled Conda installation. + +\item Generally, by default Conda relies heavily on the operating system and does not include core analysis components like \inlinecode{mkdir}, \inlinecode{ls} or \inlinecode{cp}. + Although they are generally the same between different Unix-like operating systems, they have their differences. + For example \inlinecode{mkdir -p} is a common way to build directories, but this option is only available with GNU Coreutils (default on GNU/Linux systems). + Running the same command within a Conda environment on a macOS for example, will crash. + Important packages like GNU Coreutils are available in channels like conda-forge, but they are not the default. + Therefore, many users may not recognize this, and failing to account for it, will cause unexpected crashes. + +\item Many major Conda packaging ``channels'' (for example the core Anaconda channel, or very popular conda-forge channel) do not include the C library, that a package was built with, as a dependency. + They rely on the host operating system's C library. + C is the core language of modern operating systems and even higher-level languages like Python or R are written in it, and need it to run. + Therefore if the host operating system's C library is different from the C library that a package was built with, a Conda-packaged program will crash and the project will not be executable. + Theoretically, it is possible to define a new Conda ``channel'' which includes the C library as a dependency of its software packages, but it will take too much time for any individual team to practically implement all their necessary packages, up to their high-level science software. + +\item Conda does allow a package to depend on a special build of its prerequisites (specified by a checksum, fixing its version and the version of its dependencies). + However, this is rarely practiced in the main Git repositories of channels like Anaconda and conda-forge: only the name of the high-level prerequisite packages is listed in a package's \inlinecode{meta.yaml} file, which is version-controlled. + Therefore two builds of the package from the same Git repository will result in different tarballs (depending on what prerequisites were present at build time). + In the Conda tarball (that contains the binaries and is not under version control) \inlinecode{meta.yaml} does include the exact versions of most build-time dependencies. + However, because the different software of one project may have been built at different times, if they depend on different versions of a single software there will be a conflict and the tarball cannot be rebuilt, or the project cannot be run. +\end{itemize} + +As reviewed above, the low-level dependence of Conda on the host operating system's components and build-time conditions, is the primary reason that it is very fast to install (thus making it an attractive tool to software developers who just need to reproduce a bug in a few minutes). +However, these same factors are major caveats in a scientific scenario, where long-term archivability, readability or usability are important. % alternative to `archivability`? + + +\subsubsection{Spack} +Spack is a package manager that is also influenced by Nix (similar to GNU Guix), see \citeappendix{gamblin15}. + But unlike Nix or GNU Guix, it does not aim for full, bit-wise reproducibility and can be built without root access in any generic location. + It relies on the host operating system for the C library. + + Spack is fully written in Python, where each software package is an instance of a class, which defines how it should be downloaded, configured, built and installed. + Therefore if the proper version of Python is not present, Spack cannot be used and when incompatibilities arise in future versions of Python (similar to how Python 3 is not compatible with Python 2), software building recipes, or the whole system, have to be upgraded. + Because of such bootstrapping problems (for example how Spack needs Python to build Python and other software), it is generally a good practice to use simpler, lower-level languages/systems for a low-level operation like package management. + + +In conclusion for all package managers, there are two common issues regarding generic package managers that hinders their usage for high-level scientific projects, as listed below: +\begin{itemize} +\item {\bf\small Pre-compiled/binary downloads:} Most package managers (excluding Nix or its derivatives) only download the software in a binary (pre-compiled) format. + This allows users to download it very fast and almost instantaneously be able to run it. + However, to provide for this, servers need to keep binary files for each build of the software on different operating systems (for example Conda needs to keep binaries for Windows, macOS and GNU/Linux operating systems). + It is also necessary for them to store binaries for each build, which includes different versions of its dependencies. + Maintaining such a large binary library is expensive, therefore once the shelf-life of a binary has expired, it will be removed, causing problems for projects that depends on them. + +\item {\bf\small Adding high-level software:} Packaging new software is not trivial and needs a good level of knowledge/experience with that package manager. +For example each has its own special syntax/standards/languages, with pre-defined variables that must already be known before someone can packaging new software for them. + +However, in many research projects, the most high-level analysis software are written by the team that is doing the research, and they are its primary users, even when the software are distributed with free licenses on open repositories. +Although active package manager members are commonly very supportive in helping to package new software, many teams may not be able to make that extra effort/time investment. +As a result, they manually install their high-level software in an uncontrolled, or non-standard way, thus jeopardizing the reproducibility of the whole work. +This is another consequence of detachment of the package manager from the project doing the analysis. +\end{itemize} + +Addressing these issues has been the basic reason d'\^etre of the proposed criteria: based on the completeness criteria, instructions to download and build the packages are included within the actual science project and no special/new syntax/language is used: software download, building and installation is done with the same language/syntax that researchers manage their research: using the shell (by default GNU Bash in Maneage) and Make (by default, GNU Make in Maneage). + + + +\subsection{Version control} +\label{appendix:versioncontrol} +A scientific project is not written in a day; it usually takes more than a year. +During this time, the project evolves significantly from its first starting date and components are added or updated constantly as it approaches completion. +Added with the complexity of modern computational projects, is not trivial to manually track this evolution, and the evolution's affect of on the final output: files produced in one stage of the project can mistakenly be used by an evolved analysis environment in later stages (where the project has evolved). + +Furthermore, scientific projects do not progress linearly: earlier stages of the analysis are often modified after later stages are written. +This is a natural consequence of the scientific method; where progress is defined by experimentation and modification of hypotheses (results from earlier phases). + +It is thus very important for the integrity of a scientific project that the state/version of its processing is recorded as the project evolves. +For example better methods are found or more data arrive. +Any intermediate dataset that is produced should also be tagged with the version of the project at the time it was created. +In this way, later processing stages can make sure that they can safely be used, i.e., no change has been made in their processing steps. + +Solutions to keep track of a project's history have existed since the early days of software engineering in the 1970s and they have constantly improved over the last decades. +Today the distributed model of ``version control'' is the most common, where the full history of the project is stored locally on different systems and can easily be integrated. +There are many existing version control solutions, for example CVS, SVN, Mercurial, GNU Bazaar, or GNU Arch. +However, currently, Git is by far the most commonly used in individual projects. +Git is also the foundation on which this paper's proof of concept (Maneage) is built upon. +Archival systems aiming for long term preservation of software like Software Heritage \citeappendix{dicosmo18} are also modeled on Git. +Hence we will just review Git here, but the general concept of version control is the same in all implementations. + +\subsubsection{Git} +With Git, changes in a project's contents are accurately identified by comparing them with their previous version in the archived Git repository. +When the user decides the changes are significant compared to the archived state, they can ``commit'' the changes into the history/repository. +The commit involves copying the changed files into the repository and calculating a 40 character checksum/hash that is calculated from the files, an accompanying ``message'' (a narrative description of the purpose/goals of the changes), and the previous commit (thus creating a ``chain'' of commits that are strongly connected to each other like Figure \ref{fig:branching}). +For example \inlinecode{f4953cc\-f1ca8a\-33616ad\-602ddf\-4cd189\-c2eff97b} is a commit identifier in the Git history of this project. +Commits are is commonly summarized by the checksum's first few characters, for example \inlinecode{f4953cc}. + +With Git, making parallel ``branches'' (in the project's history) is very easy and its distributed nature greatly helps in the parallel development of a project by a team. +The team can host the Git history on a webpage and collaborate through that. +There are several Git hosting services for example \href{http://codeberg.org}{codeberg.org}, \href{http://gitlab.com}{gitlab.com}, \href{http://bitbucket.org}{bitbucket.org} or \href{http://github.com}{github.com} (among many others). +Storing the changes in binary files is also possible in Git, however it is most useful for human-readable plain-text sources. + + + +\subsection{Job management} +\label{appendix:jobmanagement} +Any analysis will involve more than one logical step. +For example it is first necessary to download a dataset and do some preparations on it before applying the research software on it, and finally to make visualizations/tables that can be imported into the final report. +Each one of these is a logically independent step, which needs to be run before/after the others in a specific order. + +Hence job management is a critical component of a research project. +There are many tools for managing the sequence of jobs, below we review the most common ones that are also used the existing reproducibility solutions of Appendix \ref{appendix:existingsolutions}. + +\subsubsection{Manual operation with narrative} +\label{appendix:manual} +The most commonly used workflow system for many researchers is to run the commands, experiment on them and keep the output when they are happy with it. +As an improvement, some also keep a narrative description of what they ran. +Atleast in our personal experience with colleagues, this method is still being heavily practiced by many researchers. +Given that many researchers do not get trained well in computational methods, this is not surprizing and as discussed in Section \ref{discussion}, we believe that improved literacy in computational methods is the single most important factor for the integrity/reproducibility of modern science. + +\subsubsection{Scripts} +\label{appendix:scripts} +Scripts (in any language, for example GNU Bash, or Python) are the most common ways of organizing a series of steps. +They are primarily designed to execute each step sequentially (one after another), making them also very intuitive. +However, as the series of operations become complex and large, managing the workflow in a script will become highly complex. + +For example if 90\% of a long project is already done and a researcher wants to add a followup step, a script will go through all the previous steps (which can take significant time). +In other scenarios, when a small step in the middle of an analysis has to be changed, the full analysis needs to be re-run from the start. +Scripts have no concept of dependencies, forcing authors to ``temporarily'' comment parts of that they do not want to be re-run (forgetting to un-comment such parts are the most common cause of frustration for the authors and others attempting to reproduce the result). + +Such factors discourage experimentation, which is a critical component of the scientific method. +It is possible to manually add conditionals all over the script to add dependencies or only run certain steps at certain times, but they just make it harder to read, and introduce many bugs themselves. +Parallelization is another drawback of using scripts. +While its not impossible, because of the high-level nature of scripts, it is not trivial and parallelization can also be very inefficient or buggy. + + +\subsubsection{Make} +\label{appendix:make} +Make was originally designed to address the problems mentioned above for scripts \citeappendix{feldman79}. +In particular, it addresses the context of managing the compilation of software programs that involve many source code files. +With Make, the source files of a program that have not been changed are not recompiled. +Moreover, when two source files do not depend on each other, and both need to be rebuilt, they can be built in parallel. +This was found to greatly help in debugging software projects, and in speeding up test builds, giving Make a core place in software development over the last 40 years. + +The most common implementation of Make, since the early 1990s, is GNU Make. +Make was also the framework used in the first attempts at reproducible scientific papers \cite{claerbout1992,schwab2000}. +Our proof-of-concept (Maneage) also uses Make to organize its workflow. +Here, we complement that section with more technical details on Make. + +Usually, the top-level Make instructions are placed in a file called Makefile, but it is also common to use the \inlinecode{.mk} suffix for custom file names. +Each stage/step in the analysis is defined through a \emph{rule}. +Rules define \emph{recipes} to build \emph{targets} from \emph{pre-requisites}. +In \new{Unix-like operating systems}, everything is a file, even directories and devices. +Therefore all three components in a rule must be files on the running filesystem. + +To decide which operation should be re-done when executed, Make compares the time stamp of the targets and prerequisites. +When any of the prerequisite(s) is newer than a target, the recipe is re-run to re-build the target. +When all the prerequisites are older than the target, that target does not need to be rebuilt. +The recipe can contain any number of commands, they should just all start with a \inlinecode{TAB}. +Going deeper into the syntax of Make is beyond the scope of this paper, but we recommend interested readers to consult the GNU Make manual for a nice introduction\footnote{\inlinecode{\url{http://www.gnu.org/software/make/manual/make.pdf}}}. + +\subsubsection{Snakemake} +is a Python-based workflow management system, inspired by GNU Make (which is the job organizer in Maneage), that is aimed at reproducible and scalable data analysis \citeappendix{koster12}\footnote{\inlinecode{\url{https://snakemake.readthedocs.io/en/stable}}}. +It defines its own language to implement the ``rule'' concept in Make within Python. +Currently it requires Python 3.5 (released in September 2015) and above, while Snakemake was originally introduced in 2012. +Hence it is not clear if older Snakemake source files can be executed today. +This as reviewed in many tools here, this is a major longevity problem when using highlevel tools as the skeleton of the workflow. +Technically, calling commond-line programs within Python is very slow and using complex shell scripts in each step will involve a lot quotations that make the code hard to read. + +\subsubsection{Bazel} +Bazel\footnote{\inlinecode{\url{https://bazel.build}}} is a high-level job organizer that depends on Java and Python and is primarily tailored to software developers (with features like facilitating linking of libraries through its high level constructs). + +\subsubsection{SCons} +\label{appendix:scons} +Scons is a Python package for managing operations outside of Python (in contrast to CGAT-core, discussed below, which only organizes Python functions). +In many aspects it is similar to Make, for example it is managed through a `SConstruct' file. +Like a Makefile, SConstruct is also declarative: the running order is not necessarily the top-to-bottom order of the written operations within the file (unlike the imperative paradigm which is common in languages like C, Python, or FORTRAN). +However, unlike Make, SCons does not use the file modification date to decide if it should be remade. +SCons keeps the MD5 hash of all the files (in a hidden binary file) to check if the contents have changed. + +SCons thus attempts to work on a declarative file with an imperative language (Python). +It also goes beyond raw job management and attempts to extract information from within the files (for example to identify the libraries that must be linked while compiling a program). +SCons is therefore more complex than Make and its manual is almost double that of GNU Make. +Besides added complexity, all these ``smart'' features decrease its performance, especially as files get larger and more numerous: on every call, every file's checksum has to be calculated, and a Python system call has to be made (which is computationally expensive). + +Finally, it has the same drawback as any other tool that uses high-level languages, see Section \ref{appendix:highlevelinworkflow}. +We encountered such a problem while testing SCons: on the Debian-10 testing system, the \inlinecode{python} program pointed to Python 2. +However, since Python 2 is now obsolete, SCons was built with Python 3 and our first run crashed. +To fix it, we had to either manually change the core operating system path, or the SCons source hashbang. +The former will conflict with other system tools that assume \inlinecode{python} points to Python-2, the latter may need root permissions for some systems. +This can also be problematic when a Python analysis library, may require a Python version that conflicts with the running SCons. + +\subsubsection{CGAT-core} +CGAT-Core is a Python package for managing workflows, see \citeappendix{cribbs19}. +It wraps analysis steps in Python functions and uses Python decorators to track the dependencies between tasks. +It is used papers like \citeappendix{jones19}, but as mentioned in \citeappendix{jones19} it is good for managing individual outputs (for example separate figures/tables in the paper, when they are fully created within Python). +Because it is primarily designed for Python tasks, managing a full workflow (which includes many more components, written in other languages) is not trivial in it. +Another drawback with this workflow manager is that Python is a very high-level language where future versions of the language may no longer be compatible with Python 3, that CGAT-core is implemented in (similar to how Python 2 programs are not compatible with Python 3). + +\subsubsection{Guix Workflow Language (GWL)} +GWL is based on the declarative language that GNU Guix uses for package management (see Appendix \ref{appendix:packagemanagement}), which is itself based on the general purpose Scheme language. +It is closely linked with GNU Guix and can even install the necessary software needed for each individual process. +Hence in the GWL paradigm, software installation and usage does not have to be separated. +GWL has two high-level concepts called ``processes'' and ``workflows'' where the latter defines how multiple processes should be executed together. + +In conclusion, shell scripts and Make are very common and extensively used by users of Unix-based OSs (which are most commonly used for computations). +They have also existed for several decades and are robust and mature. +Many researchers are also already familiar with them and have already used them. +As we see in this appendix, the list of necessary tools for the various stages of a research project (an independent environment, package managers, job organizers, analysis languages, writing formats, editors and etc) is already very large. +Each software has its own learning curve, which is a heavy burden for a natural or social scientist for example. +Most other workflow management tools are yet another language that have to be mastered. + +Furthermore, high-level and specific solutions will evolve very fast causing disruptions in the reproducible framework. +A good example is Popper \citeappendix{jimenez17} which initially organized its workflow through the HashiCorp configuration language (HCL) because it was the default in GitHub. +However, in September 2019, GitHub dropped HCL as its default configuration language, so Popper is now using its own custom YAML-based workflow language, see Appendix \ref{appendix:popper} for more on Popper. + +\subsubsection{Nextflow (2013)} +Nextflow\footnote{\inlinecode{\url{https://www.nextflow.io}}} \citeappendix{tommaso17} workflow language with a command-line interface that is written in Java. + +\subsubsection{Generic workflow specifications (CWL and WDL)} +Due to the variety of custom workflows used in existing reproducibility solution (like those of Appendix \ref{appendix:existingsolutions}), some attempts have been made to define common workflow standards like the Common workflow language (CWL\footnote{\inlinecode{\url{https://www.commonwl.org}}}, with roots in Make, formatted in YAML or JSON) and Workflow Description Language (WDL\footnote{\inlinecode{\url{https://openwdl.org}}}, formatted in JSON). +These are primarily specifications/standards rather than software, so ideally translators can be written between the various workflow systems to make them more interoperable. + + +\subsection{Editing steps and viewing results} +\label{appendix:editors} +In order to later reproduce a project, the analysis steps must be stored in files. +For example Shell, Python or R scripts, Makefiles, Dockerfiles, or even the source files of compiled languages like C or FORTRAN. +Given that a scientific project does not evolve linearly and many edits are needed as it evolves, it is important to be able to actively test the analysis steps while writing the project's source files. +Here we review some common methods that are currently used. + +\subsubsection{Text editors} +The most basic way to edit text files is through simple text editors which just allow viewing and editing such files, for example \inlinecode{gedit} on the GNOME graphic user interface. +However, working with simple plain text editors like \inlinecode{gedit} can be very frustrating since its necessary to save the file, then go to a terminal emulator and execute the source files. +To solve this problem there are advanced text editors like GNU Emacs that allow direct execution of the script, or access to a terminal within the text editor. +However, editors that can execute or debug the source (like GNU Emacs), just run external programs for these jobs (for example GNU GCC, or GNU GDB), just as if those programs was called from outside the editor. + +With text editors, the final edited file is independent of the actual editor and can be further edited with another editor, or executed without it. +This is a very important feature that is not commonly present for other solutions mentioned below. +Another very important advantage of advanced text editors like GNU Emacs or Vi(m) is that they can also be run without a graphic user interface, directly on the command-line. +This feature is critical when working on remote systems, in particular high performance computing (HPC) facilities that do not provide a graphic user interface. +Also, the commonly used minimalistic containers do not include a graphic user interface. + +\subsubsection{Integrated Development Environments (IDEs)} +To facilitate the development of source files, IDEs add software building and running environments as well as debugging tools to a plain text editor. +Many IDEs have their own compilers and debuggers, hence source files that are maintained in IDEs are not necessarily usable/portable on other systems. +Furthermore, they usually require a graphic user interface to run. +In summary IDEs are generally very specialized tools, for special projects and are not a good solution when portability (the ability to run on different systems and at different times) is required. + +\subsubsection{Jupyter} +\label{appendix:jupyter} +Jupyter (initially IPython) \citeappendix{kluyver16} is an implementation of Literate Programming \citeappendix{knuth84}. +The main user interface is a web-based ``notebook'' that contains blobs of executable code and narrative. +Jupyter uses the custom built \inlinecode{.ipynb} format\footnote{\inlinecode{\url{https://nbformat.readthedocs.io/en/latest}}}. +Jupyter's name is a combination of the three main languages it was designed for: Julia, Python and R. +The \inlinecode{.ipynb} format, is a simple, human-readable (can be opened in a plain-text editor) file, formatted in JavaScript Object Notation (JSON). +It contains various kinds of ``cells'', or blobs, that can contain narrative description, code, or multi-media visualizations (for example images/plots), that are all stored in one file. +The cells can have any order, allowing the creation of a literal programming style graphical implementation, where narrative descriptions and executable patches of code can be intertwined. +For example to have a paragraph of text about a patch of code, and run that patch immediately in the same page. + +The \inlinecode{.ipynb} format does theoretically allow dependency tracking between cells, see IPython mailing list (discussion started by Gabriel Becker from July 2013\footnote{\url{https://mail.python.org/pipermail/ipython-dev/2013-July/010725.html}}). +Defining dependencies between the cells can allow non-linear execution which is critical for large scale (thousands of files) and complex (many dependencies between the cells) operations. +It allows automation, run-time optimization (deciding not to run a cell if its not necessary) and parallelization. +However, Jupyter currently only supports a linear run of the cells: always from the start to the end. +It is possible to manually execute only one cell, but the previous/next cells that may depend on it, also have to be manually run (a common source of human error, and frustration for complex operations). +Integration of directional graph features (dependencies between the cells) into Jupyter has been discussed, but as of this publication, there is no plan to implement it (see Jupyter's GitHub issue 1175\footnote{\inlinecode{\url{https://github.com/jupyter/notebook/issues/1175}}}). + +The fact that the \inlinecode{.ipynb} format stores narrative text, code and multi-media visualization of the outputs in one file, is another major hurdle: +The files can easy become very large (in volume/bytes) and hard to read. +Both are critical for scientific processing, especially the latter: when a web browser with proper JavaScript features is not available (can happen in a few years). +This is further exacerbated by the fact that binary data (for example images) are not directly supported in JSON and have to be converted into much less memory-efficient textual encodings. + +Finally, Jupyter has an extremely complex dependency graph: on a clean Debian 10 system, Pip (a Python package manager that is necessary for installing Jupyter) required 19 dependencies to install, and installing Jupyter within Pip needed 41 dependencies! +\citeappendix{hinsen15} reported such conflicts when building Jupyter into the Active Papers framework (see Appendix \ref{appendix:activepapers}). +However, the dependencies above are only on the server-side. +Since Jupyter is a web-based system, it requires many dependencies on the viewing/running browser also (for example special JavaScript or HTML5 features, which evolve very fast). +As discussed in Appendix \ref{appendix:highlevelinworkflow} having so many dependencies is a major caveat for any system regarding scientific/long-term reproducibility (as opposed to industrial/immediate reproducibility). +In summary, Jupyter is most useful in manual, interactive and graphical operations for temporary operations (for example educational tutorials). + + + + + + +\subsection{Project management in high-level languages} +\label{appendix:highlevelinworkflow} + +Currently the most popular high-level data analysis language is Python. +R is closely tracking it, and has superseded Python in some fields, while Julia \citeappendix{bezanson17} is quickly gaining ground. +These languages have themselves superseded previously popular languages for data analysis of the previous decades, for example Java, Perl or C++. +All are part of the C-family programming languages. +In many cases, this means that the tools to use that language are written in C, which is the language of modern operating systems. + +Scientists, or data analysts, mostly use these higher-level languages. +Therefore they are naturally drawn to also apply the higher-level languages for lower-level project management, or designing the various stages of their workflow. +For example Conda or Spack (Appendix \ref{appendix:packagemanagement}), CGAT-core (Appendix \ref{appendix:jobmanagement}), Jupyter (Appendix \ref{appendix:editors}) or Popper (Appendix \ref{appendix:popper}) are written in Python. +The discussion below applies to both the actual analysis software and project management software. +In this context, its more focused on the latter. + +Because of their nature, higher-level languages evolve very fast, creating incompatibilities on the way. +The most prominent example is the transition from Python 2 (released in 2000) to Python 3 (released in 2008). +Python 3 was incompatible with Python 2 and it was decided to abandon the former by 2015. +However, due to community pressure, this was delayed to January 1st, 2020. +The end-of-life of Python 2 caused many problems for projects that had invested heavily in Python 2: all their previous work had to be translated, for example see \citeappendix{jenness17} or Appendix \ref{appendix:sciunit}. +Some projects could not make this investment and their developers decided to stop maintaining it, for example VisTrails (see Appendix \ref{appendix:vistrails}). + +The problems were not just limited to translation. +Python 2 was still actively being actively used during the transition period (and is still being used by some, after its end-of-life). +Therefore, developers of packages used by others had to maintain (for example fix bugs in) both versions in one package. +This is not particular to Python, a similar evolution occurred in Perl: in 2000 it was decided to improve Perl 5, but the proposed Perl 6 was incompatible with it. +However, the Perl community decided not to abandon Perl 5, and Perl 6 was eventually defined as a new language that is now officially called ``Raku'' (\url{https://raku.org}). + +It is unreasonably optimistic to assume that high-level languages will not undergo similar incompatible evolutions in the (not too distant) future. +For software developers, this is not a problem at all: non-scientific software, and the general population's usage of them, has a similarly fast evolvution. +Hence, it is rarely (if ever) necessary to look into codes that are more than a couple of years old. +However, in the sciences (which are commonly funded by public money) this is a major caveat for the longer-term usability of solutions that are designed in such high level languages. + +In summary, in this section we are discussing the bootstrapping problem as regards scientific projects: the workflow/pipeline can reproduce the analysis and its dependencies, but the dependencies of the workflow itself cannot not be ignored. +Beyond technical, low-level, problems for the developers mentioned above, this causes major problems for scientific project management as listed below: + +\subsubsection{Dependency hell} +The evolution of high-level languages is extremely fast, even within one version. +For example packages that are written in Python 3 often only work with a special interval of Python 3 versions (for example newer than Python 3.6). +This is not just limited to the core language, much faster changes occur in their higher-level libraries. +For example version 1.9 of Numpy (Python's numerical analysis module) discontinued support for Numpy's predecessor (called Numeric), causing many problems for scientific users \citeappendix{hinsen15}. + +On the other hand, the dependency graph of tools written in high-level languages is often extremely complex. +For example see Figure 1 of \cite{alliez19}, it shows the dependencies and their inter-dependencies for Matplotlib (a popular plotting module in Python). +Acceptable version intervals between the dependencies will cause incompatibilities in a year or two, when a robust package manager is not used (see Appendix \ref{appendix:packagemanagement}). + +Since a domain scientist does not always have the resources/knowledge to modify the conflicting part(s), many are forced to create complex environments with different versions of Python and pass the data between them (for example just to use the work of a previous PhD student in the team). +This greatly increases the complexity of the project, even for the principal author. +A good reproducible workflow can account for these different versions. +However, when the actual workflow system (not the analysis software) is written in a high-level language this will cause a major problem. + +For example, merely installing the Python installer (\inlinecode{pip}) on a Debian system (with \inlinecode{apt install pip2} for Python 2 packages), required 32 other packages as dependencies. +\inlinecode{pip} is necessary to install Popper and Sciunit (Appendices \ref{appendix:popper} and \ref{appendix:sciunit}). +As of this writing, the \inlinecode{pip3 install popper} and \inlinecode{pip2 install sciunit2} commands for installing each, required 17 and 26 Python modules as dependencies. +It is impossible to run either of these solutions if there is a single conflict in this very complex dependency graph. +This problem actually occurred while we were testing Sciunit: even though it installed, it could not run because of conflicts (its last commit was only 1.5 years old), for more see Appendix \ref{appendix:sciunit}. +\citeappendix{hinsen15} also report a similar problem when attempting to install Jupyter (see Appendix \ref{appendix:editors}). +Of course, this also applies to tools that these systems use, for example Conda (which is also written in Python, see Appendix \ref{appendix:packagemanagement}). + + + + + +\subsubsection{Generational gap} +This occurs primarily for domain scientists (for example astronomers, biologists or social sciences). +Once they have mastered one version of a language (mostly in the early stages of their career), they tend to ignore newer versions/languages. +The inertia of programming languages is very strong. +This is natural, because they have their own science field to focus on, and re-writing their high-level analysis toolkits (which they have curated over their career and is often only readable/usable by themselves) in newer languages every few years requires too much investment and time. + +When this investment is not possible, either the mentee has to use the mentor's old method (and miss out on all the new tools, which they need for the future job prospects), or the mentor has to avoid implementation details in discussions with the mentee, because they do not share a common language. +The authors of this paper have personal experiences in both mentor/mentee relational scenarios. +This failure to communicate in the details is a very serious problem, leading to the loss of valuable inter-generational experience. diff --git a/tex/src/appendix-necessity.tex b/tex/src/appendix-necessity.tex new file mode 100644 index 0000000..6578e5f --- /dev/null +++ b/tex/src/appendix-necessity.tex @@ -0,0 +1,79 @@ + + +\section{Necessity for reproducible research} +\label{appendix:necessity} + +The increasing volume and complexity of data analysis has been highly productive, giving rise to a new branch of ``Big Data'' in many fields of the sciences and industry. +However, given its inherent complexity, the mere results are barely useful alone. +Questions such as these commonly follow any such result: +What inputs were used? +What operations were done on those inputs? How were the configurations or training data chosen? +How did the quantitative results get visualized into the final demonstration plots, figures or narrative/qualitative interpretation? +Could there be a bias in the visualization? +See Figure \ref{fig:questions} for a more detailed visual representation of such questions for various stages of the workflow. + +In data science and database management, this type of metadata is commonly known as \emph{data provenance} or \emph{data lineage}. +Data lineage is being increasingly demanded for integrity checking from both the scientific, industrial and legal domains. +Notable examples in each domain are respectively the ``Reproducibility crisis'' in the sciences that was reported by \emph{Nature} after a large survey \citeappendix{baker16}, and the General Data Protection Regulation (GDPR) by the European Parliament and the California Consumer Privacy Act (CCPA), implemented in 2018 and 2020, respectively. +The former argues that reproducibility (as a test on sufficiently conveying the data lineage) is necessary for other scientists to study, check and build-upon each other's work. +The latter requires the data intensive industry to give individual users control over their data, effectively requiring thorough management and knowledge of the data lineage. +Besides regulation and integrity checks, having robust data governance (management of data lineage) in a project can be very productive: it enables easy debugging, experimentation on alternative methods, or optimization of the workflow. + +In the sciences, the results of a project's analysis are published as scientific papers, which have traditionally been the primary conveyor of the lineage of the results: usually in narrative form, especially within the ``Methods'' section of the paper. +From our own experience, this section is often that which is the most intensively discussed during peer review and conference presentations, showing its importance. +After all, a result is defined as ``scientific'' based on its \emph{method} (the ``scientific method''), or lineage in data-science terminology. +In industry, however, data governance is usually kept as a trade secret and is not published openly or widely scrutinized. +Therefore, the main practical focus here will be in the scientific front, which has traditionally been more open to the publication of methods and anonymous peer scrutiny. + +\begin{figure*}[t] + \begin{center} + \includetikz{figure-project-outline}{width=\linewidth} + \end{center} + \caption{\label{fig:questions}Graph of a generic project's workflow (connected through arrows), highlighting the various issues and questions on each step. + The green boxes with sharp edges are inputs and the blue boxes with rounded corners are the intermediate or final outputs. + The red boxes with dashed edges highlight the main questions at various stages in the workchain. + The orange box, surrounding the software download and build phases, lists some commonly recognized solutions to the questions in it; for morediscussion, see Appendix \ref{appendix:independentenvironment}. + } +\end{figure*} + +The traditional format of a scientific paper has been very successful in conveying the method and the results during recent centuries. +However, the complexity mentioned above has made it impossible to describe all the analytical steps of most modern projects to a sufficient level of detail. +Citing this difficulty, many authors limit themselves to describing the very high-level generalities of their analysis, while even the most basic calculations (such as the mean of a distribution) can depend on the software implementation. + +Due to the complexity of modern scientific analysis, a small deviation in some of the different steps involved can lead to significant differences in the final result. +Publishing the precise codes of the analysis is the only guarantee of allowing this to be investigated. +For example, \citeappendix{smart18} describes how a 7-year old conflict in theoretical condensed matter physics was only identified after the different groups' codes were shared. +Nature is already a black box that we are trying hard to unlock and understand. +Not being able to experiment on the methods of other researchers is an artificial and self-imposed black box, wrapped over the original, and wasting much of researchers' time and energy. + +A dramatic example showing the importance of sharing code is \citeappendix{miller06}, in which a mistaken flipping of a column was discovered, leading to the retraction of five papers in major journals, including \emph{Science}. +Ref.\/ \citeappendix{baggerly09} highlighted the inadequate narrative description of analysis in several papers and showed the prevalence of simple errors in published results, ultimately calling their work ``forensic bioinformatics''. +References \citeappendix{herndon14} and \citeappendix{horvath15} also reported similar situations and \citeappendix{ziemann16} concluded that one-fifth of papers with supplementary Microsoft Excel gene lists contain erroneous gene name conversions. +Such integrity checks are a critical component of the scientific method, but are only possible with access to the data and codes and \emph{cannot be resolved from analysing the published paper alone}. + +The completeness of a paper's published metadata (or ``Methods'' section) can be measured by a simple question: given the same input datasets (supposedly on a third-party database like \href{http://zenodo.org}{zenodo.org}), can another researcher reproduce the exact same result automatically, without needing to contact the authors? +Several studies have attempted to answer this with different levels of detail. +For example \citeappendix{allen18} found that roughly half of the papers in astrophysics do not even mention the names of any analysis software they have used, while \cite{menke20} found that the fraction of papers explicitly mentioning their software tools has greatly improved in medical journals over the last two decades. + +Ref.\/ \citeappendix{ioannidis2009} attempted to reproduce 18 published results by two independent groups, but only fully succeeded in two of them and partially in six. +Ref.\/ \citeappendix{chang15} attempted to reproduce 67 papers in well-regarded economic journals with data and code: only 22 could be reproduced without contacting authors, and more than half could not be replicated at all. +Ref.\/ \citeappendix{stodden18} attempted to replicate the results of 204 scientific papers published in the journal Science \emph{after} that journal adopted a policy of publishing the data and code associated with the papers. +Even though the authors were contacted, the success rate was $26\%$. +Generally, this problem is unambiguously felt in the community: \citeappendix{baker16} surveyed 1574 researchers and found that only $3\%$ did not see a ``reproducibility crisis''. + +This is not a new problem in the sciences: in 2011, Elsevier conducted an ``Executable Paper Grand Challenge'' \citeappendix{gabriel11}. +The proposed solutions were published in a special edition. +Some of them are reviewed in Appendix \ref{appendix:existingsolutions}, but most have not been continued since then. +In 2005, Ref.\/ \citeappendix{ioannidis05} argued that ``most claimed research findings are false''. +Even earlier, in the 1990s, Refs \cite{schwab2000}, \citeappendix{buckheit1995} and \cite{claerbout1992} described this same problem very eloquently and provided some of the solutions that they adopted. +While the situation has improved since the early 1990s, the problems mentioned in these papers will resonate strongly with the frustrations of today's scientists. +Even earlier yet, through his famous quartet, Anscombe \citeappendix{anscombe73} qualitatively showed how the distancing of researchers from the intricacies of algorithms and methods can lead to misinterpretation of the results. +One of the earliest such efforts we found was \citeappendix{roberts69}, who discussed conventions in FORTRAN programming and documentation to help in publishing research codes. + +From a practical point of view, for those who publish the data lineage, a major problem is the fast evolving and diverse software technologies and methodologies that are used by different teams in different epochs. +Ref.\/ \citeappendix{zhao12} describes it as ``workflow decay'' and recommends preserving these auxiliary resources. +But in the case of software, this is not as straightforward as for data: if preserved in binary form, software can only be run on a certain operating system on particular hardware, and if kept as source code, its build dependencies and build configuration must also be preserved. +Ref.\/ \citeappendix{gronenschild12} specifically studies the effect of software version and environment and encourages researchers to not update their software environment. +However, this is not a practical solution because software updates are necessary, at least to fix bugs in the same research software. +Generally, software is not a secular component of projects, where one software package can easily be swapped with another. +Projects are built around specific software technologies, and research in software methods and implementations is itself a vibrant research topic in many domains \citeappendix{dicosmo19}. |