The Experimental Development Team (Xdev)

The Experimental Development Team (Xdev) was born out of Pangeo. Matthew Long and I felt that Pangeo was a movement that NCAR needed to join. Not only does Pangeo offer a roadmap for interactive “Big Data” analysis for climate and weather scientists, but it demonstrates a revolutionary culture of collaboration between scientists and software engineers spanning both academia and industry. Pangeo realizes the dream of open source software and open development that so many people evangelize but so few achieve. But to realize this dream at NCAR, we need to do more than just use Pangeo Environments. We need to do more than just teach scientists how to use Xarray and Dask and Jupyter, or how program in Python, or even how to develop an NCL replacement in Python that leverages the Pangeo stack.

NCAR needs to change how its scientists and software engineers relate to one another, and how technology for data science and analysis at NCAR is developed openly.

Xdev was created to be the catalyst for this cultural change at NCAR, not just by teaching Python but by working with scientists in direct collaboration on software, and therefore teaching open development by example.

To achieve this, Xdev acts as navigators for NCAR scientists in the Pangeo sea. As such, Xdev helps NCAR scientists adopt a common approach to enabling broad multidisciplinary collaboration through its education and development efforts. Xdev’s education efforts to teach scientists at NCAR (and beyond) Python and the Pangeo stack help build and strengthen the Pangeo community of technology users. Meanwhile, Xdev’s development efforts to collaborate with scientists on Pangeo technology development help teach open development and open source software best practices through example.

The Xdev Manifesto

Xdev is an NCAR entity, and therefore must satisfy NCAR priorities and strategic objectives. Xdev is also committed to the external Pangeo community, and therefore it behooves us to enumerate our technical values to highlight how Xdev’s values align with Pangeo.

  1. We value interactivity over batch processing. While batch processing is still necessary for many things, we believe that the need for curiosity-driven analysis is greater and that this should be the priority.

  2. We value open source over closed source solutions. While some closed source solutions may be better for specific needs, especially if a monetary investment has already been made, we believe that open source solutions are a better long term solution in general.

  3. We value scalability. While serial, non-scalable solutions can be reasonable algorithm prototypes, they are of limited value in the era of Big Data. We believe that scalability should be considered at all times.

  4. We value workflow portability. While having solutions that only work on NCAR systems can solve immediate problems, we want to reduce the burden on scientists for having to learn different solutions on different platforms, whether they be HPC or cloud. Scientists should decide where to do their analysis based on where the data is hosted, not based on where they can run the specific code they want to run.

  5. We value transparency over opacity. While some work for many reasons cannot be made public, we choose to do all of our work in the open so that others can learn from us, leverage our work, and work with us at any time.

  6. We value reproducibility. While static, non-reproducible scientific communication still has value as the traditional mechanism for exchanging knowledge throughout scientific disciplines, we believe reproducible mechanisms for publishing and sharing scientific knowledge are more important and effective.

By affirming these values, we focus most of our interest on the Pangeo stack that is supported by Xarray, Dask, and Jupyter, as these technologies provide an exceptional foundation for building a replete ecosystem to support scalable analytics.

A note on reproducibility

From Xdev’s perspective, reproducibility is primarily an issue of scientific reproducibility, and not (for example) bit-for-bit reproducibility. From our perspective, reproducibility is all about making sure that research, analysis, and results can actually be reproduced by the people with whom scientific results are shared. This is why we encourage Jupyter Notebooks. Jupyter Notebooks are the basis for an “executable paper.” A number of entities are already looking into peer-reviewed notebook publication journals, and the future of notebooks as the mechanism for reproducible publications is very promising. There are, however, technical challenges to making Jupyter Notebooks truly reproducible, even in a scientific sense. A Jupyter Notebook file (i.e., a file with the .ipynb extension) will not run properly (or even at all) without the correct environment (or kernel) available in which the notebook code will run. If the notebook requires special packages to run without error, then these packages need to be installed for every user who wants to execute that notebook. Containerization is one solution to this problem, where notebooks are shared with the necessary environment in which they can run. This is the mechanism behind the Jupyter Project’s Binder utility. Binder works by sharing notebooks via publicly accessible Git repositories inside which are files defining the environment (and how to build the environment) needed by the notebooks. A BinderHub server then can be used to retrieve the notebooks and the environment definitions to build a container image from which the notebooks can be run on the cloud.

Mission & Vision

Xdev’s mission is to advance open reproducible science through interactive Big Data analysis.

We envision a community of scientists at NCAR and beyond working together to develop and maintain scientific workflows and sharable software to support those workflows on a platform where these workflows can be efficiently developed in collaboration and executed. We envision these workflows to be reproducible, using Jupyter Notebooks with fixed versioning and environments (e.g., using containerization). We envision a core group of scientists within that community who have helped develop and support the technology that makes much of this possible, who follow modern engineering best practices, and who leverage each other’s technology to quickly and efficiently develop new capabilities and new workflows, as well as help other scientists learn these skills and knowledge themselves. We envision all of this technology supporting and empowering not only NCAR staff but the entire community that NCAR supports.

As part of this vision, we also see Jupyter Notebooks becoming first-class citizens on HPC systems. In this vision, we see a different style of “analysis and visualization cluster,” one with computation occurring on HPC hardware or cloud hardware, with data held on the “edge” of the cloud, instead of copies of the data in multiple locations. We see HPC job scheduling becoming flexible and elastic, instead of only accommodating static job loads more suited for large model runs. We see containerization becoming the norm, not just to provide a pre-defined system environment for software to run, but to allow user-defined “services” on the HPC system, making it possible for scientists to share interactive analyses, instead of just static notebooks. We see a blurring of the difference between HPC and cloud computing, and in time we expect very little difference between the two at all.

As the community of scientists who use this style of analysis, and who contribute back to the software that makes this analysis possible, develops and grows, we envision Xdev’s contributions will slowly shift toward more specific engineering efforts that optimize, stabilize, and strengthen the ecosystem of technologies used by the scientific community. However, we envision Xdev still remaining a resource for scientists on all engineering matters, especially on matters that extend beyond the role that scientists nominally play. Xdev will continue to remain a resource of software engineering knowledge and expertise for the community.

Xdev & ESDS

The Earth System Data Science initiative was created to motivate scientists to build a community around using technology for open reproducible science:

The Earth System Data Science initiative aims to profoundly increase the effectiveness of the NCAR/UCAR workforce by promoting deeper collaboration centered on analytics, improving our capacity to deliver impactful, actionable, reproducible science, and serve the university community by transforming how geoscientists synthesize and extract information from large, diverse data sets.

ESDS is the force bringing the scientific community together to collaborate, and Xdev is the catalyst for cultural change. ESDS has an essentially scientific focus, while Xdev has an engineering focus. Xdev brings its expertise to bear on specific problems that scientists have with the technology. Xdev helps steer ESDS scientists toward solutions that interoperate within Pangeo Environments and helps the scientists build and depend upon a common stack of software that empowers everyone. Xdev helps teach ESDS scientists through ESDS’s training efforts and through direct collaboration with ESDS scientists on technical projects. While Xdev means to address more than just the ESDS community of scientists, ESDS is a strong partner with Xdev and an excellent source for collaboration.


Xdev is a matrixed team made of full-time and part-time software engineers. Currently, our two full-time software engineers (Anderson Banihirwe and Julia Kent) report directly to Kevin Paul. Like all matrixed teams, maintaining a clear vision of the team’s purpose and objectives across all of the team members’ supervisors is critical.

Xdev also relies closely on a group of 8 software engineers across UCAR who act as stakeholders in Xdev’s activities. These software engineers occasionally provide part-time development support on activities that directly impact their own work and provide feedback on (and direction for) Xdev’s activities, including assistance with short-term project selection, testing, and feedback, as well as tutorial input and guidance (see Governance). Most of these software engineers are also members of ESDS, but some simply represent Python expertise across UCAR and help expand and shape Xdev’s vision.

Within the core team, there are definite roles and responsibilities.

Kevin Paul is the Team Lead, and helps define and scope of development projects and education/training activities. The Team Lead seeks out and builds relationships with stakeholders across UCAR. In all of our efforts, we hope that target users will provide direct feedback to our efforts in a tight, agile iterative cycle, but we acknowledge that this cannot always be expected of scientists who are not working with the team on a regular basis. When decisions are required to keep progress moving forward, the Team Lead acts as a proxy for the user when the user is not available to provide feedback, providing functional specifications for Xdev activities from the target user perspective.

Anderson Banihirwe is the Technology Lead and helps assess the right technology to use on Xdev projects and activities. The Technical Lead is responsible for maintaining a current view of the relevant technology and assessing how processes need to change based on changing technological capabilities.

Julia Kent is the Training Lead and helps organize and maintain our tutorials and training materials, working closely with ESDS and Project Pythia. The Training Lead is responsible for keeping our educational and training mission constantly moving forward and building the knowledge base of the scientific community.

These roles have developed over time out of necessity and the people who have stepped into these roles have done so voluntarily as our activities have progressed. These roles are not set in stone, and we expect them to shift and evolve as the team moves forward.


The Xdev Team is governed by its stakeholders. In particular, there are three categories of stakeholders that govern Xdev’s activities:

  1. Member Stakeholders: Member Stakeholders are part-time members of the Xdev Team. Member Stakeholders are typically software engineers or scientists with a dedicated software engineering background and role and can provide detailed technical input on projects and other technical activities. Member Stakeholders may join the core team on short-term projects or other activities as appropriate.

  2. Community Stakeholders: Community Stakeholders consist of the broader user community, including scientists who have adopted, or are adopting, the Python language for scientific data analysis. Community Stakeholders are the experts in how Xdev’s education and technical project work should be directed, as they are the target demographic. Community Stakeholders may collaborate with the core team on short-term projects in the customer role of an Agile workflow.

  3. Leadership Stakeholders: Leadership Stakeholders include the Xdev Team Lead, the supervisors of the Xdev Members, and other leadership-level stakeholders in the success of Xdev and Xdev’s efforts. Leadership Stakeholders are responsible for a broader swath of activities within NCAR and the wider community and therefore are capable of directing collaborations between Xdev and other groups.

While Xdev activities continue, Member Stakeholders are invited to join the core team at any time in regular weekly meetings, either to provide direct feedback on short-term projects or to simply observe and stay connected. Being technical experts in their own right, Member Stakeholders are capable of being collaborators on Xdev Projects, even if they do not have the time to dedicate to project work themselves. We call them Member Stakeholders because we consider them Xdev members, and they are included in all Xdev group communications (e.g., on the mailing list). Hence, Member Stakeholders are capable of staying up-to-date with Xdev activities and providing technical feedback and guidance on a frequent basis.

Community Stakeholders are the target demographic for all Xdev activities. Therefore, Community Stakeholders are the owners of “the problem” that Xdev is trying to solve through its activities. Community Stakeholders know better than anyone else what problems they are experiencing, what they want to learn, and what they would like to accomplish scientifically via the technology that Xdev develops. Hence, Community Stakeholders have the ability to prioritize future Xdev work, propose Xdev Projects, and act as testers and the first line of feedback for Xdev results. Community Stakeholders can provide their feedback to Xdev through quarterly “Pain Point Surveys” and through a formal Xdev Project Proposal process. A list of future projects will be displayed on the Xdev website and users will be able to “upvote” projects on the list.

Leadership Stakeholders represent the institutional and community leadership, and therefore have both input on Xdev Project priorities and Xdev direction, just as Community Stakeholders do. However, Leadership Stakeholders also provide input on the workings and processes of Xdev and help address functional deficiencies, such as addressing skill gaps, assisting in grant applications, aligning Xdev activities with the institutional mission, assisting with extending Xdev across the UCAR organization and beyond. Leadership Stakeholders will meet on a quarterly, triannually, or biannually frequency, to be determined.

The frequency of these meetings and how they are conducted are expected to evolve as the team progresses.


Xdev is involved in two main activities: training and development. Xdev’s training activities center mostly around tutorials and hackathons, organized and held in collaboration with Project Pythia and ESDS. Xdev’s development activities are centered on short-term projects, one-off prototypes, and center-wide service support. These development activities focus mostly on usability issues, rather than performance enhancements. There are usability pain points that we believe need to be addressed before too much focus can be spent on optimization and performance, though we recognize that some focus on performance is necessary to assure that developed solutions are scalable or “efficient enough.”

Education & Training

Xdev’s premier product addressing education and training is the Python Tutorial Seminar Series, a bi-weekly one-hour seminar series occurring on the second and fourth Wednesday of each month at 1 PM mountain. The series started with a simple introduction to Python and has built up to Xarray and Dask. Many of these tutorials and the training material that accompanies them are being transferred to Project Pythia, with which Xdev closely collaborates. At the time of this writing, there have been 19 tutorials in the Python Seminar Series, of which 7 were prepared and conducted by non-Xdev members (mostly members from GeoCAT).

Another mechanism for training that goes beyond what can be learned simply in the classroom is the hackathon. The Pangeo community has held regular hackathons at their annual meetings, and Xdev has borrowed on the Pangeo hackathon structure. Hackathons are an excellent way for scientists who have learned the basics of the technology through tutorials and online material to build upon that knowledge through projects with measurable objectives. Hackathons are also an excellent way for scientists to learn open source culture and open development best practices, which is the primary purpose for Xdev. Many of these hackathons are held in collaboration with ESDS.

Short-Term Projects

Xdev is also planning a series of short-term projects advancing interactive Big Data analysis by focusing on pain points in the development of scientific analytic workflows. We use the term analytics, here, because the entire workflow goes well beyond simple analysis and visualization, as illustrated very coarsely in the following figure.


Fig. 3 A typical analysis workflow

This figure also attempts to illustrate how scientific workflow development is quite often iterative, such that the author of the workflow loops back to previous steps in the workflow development process many times. This means that any inefficiencies in the workflow can be magnified many times over, depending on how many times the step needs to be performed during the iterative cycle. Any of these steps can be slow and painful, and we are working toward addressing all of these steps over the next five years using the “squeaky wheel” approach: assess what is the most significant pain point, prototype a solution to alleviate this pain point, repeat.

In addition to issues around workflow development, we are concerned with reproducibility (see the note on reproducibility above). We advocate the use of Jupyter Notebooks to aid with reproducibility, but simple use of Notebooks is not enough to guarantee reproducible workflows. To ensure reproducibility, we expect to develop best practices and underlying technology to aid with workflow reproducibility, such as best practices on how to “fix” a Notebook environment, how to share Notebook environments with the Notebooks themselves, and how to version workflows and track (at least minimal) provenance of workflow components.

Xdev’s effort to catalyze cultural change at NCAR will also be a main feature of our short-term projects. Through direct collaboration with interested scientists, we hope to train a core group of scientists in the engineering and open development processes needed to help support the rest of their own community. Like hackathons, these projects serve as an excellent training mechanism for scientists to learn open source culture and open development directly through collaboration with the Xdev team. (These are the “science-software partnerships” mentioned in the ESDS Vision document.) The result is not only a publishable software technology but also newly trained scientists in the maintenance and development of the technology itself. This is vital for healthy open source software and sustainability.

As mentioned earlier, much of the scalable capabilities already exist in the Pangeo stack, but it can be cumbersome or not obvious how to construct the desired result efficiently or correctly. Education and documentation can always help address how best to use the technology, but as complex scientific workflows are constructed we observe that many scientists are doing the same things and should be able to leverage the same code without “cut-and-paste” development approaches. These projects focus on specific issues and technologies that scientists would like for their own workflows (or the workflows of their teams and groups) and how to generalize those technologies for use by a broader audience.

For more details about potential future projects and what we are currently working on, see the Short-Term Projects page.

One-Off Prototypes

In addition to short-term projects, the Xdev team also develops some in-house technologies based on observations of difficulties and struggles that we observe scientists having. Many of these technologies are “short and sweet” and do not necessitate any formal project, such as those described in the short-term project above.

One such technology is Jupyter-Forward, which was developed to give users the ability to launch an instance of JupyterLab on a remote machine via SSH. This allows users to use a specific version of JupyterLab, instead of the one supported by any existing JupyterHub, but it also allows users to use JupyterLab on a remote machine that doesn’t even have a running JupyterHub at all.

Another example of an Xdev-authored in-house technology is Intake-ESM. Intake provides a mechanism for translating queryable scientific vocabulary into data assets that can then be easily loaded into analysis notebooks as Xarray Datasets. Intake-ESM is a plugin for Intake that makes large, cumbersome datasets of hundreds or even thousands of Earth system model (ESM) output files easy to access and load.

One of our most recent in-house technologies is Xcollection, an extension of the Xarray data model to be able to handle dictionaries of Xarray Datasets. This extension allows you to apply common operations across multiple datasets easily, without needing for the datasets be merged (or mergeable) into a single dataset.

To see more about our products, visit the Xdev Packages page.

Center-Wide Service Support

NCAR, and CISL in particular, provide technological services that leverage NCAR’s computing capabilities, such as our HPC environments, the NCAR JupyterHub, and others. These services interface with a great deal of software developed by the Pangeo community and by Xdev, at the very least because the software is run in those environments and through those services. Hence, Xdev is committed to helping CISL sysadmins maintain and improve these services to create a better experience for our users.

The following diagram shows many of the elements that comprise the Pangeo Environment hosted by NCAR as part of our JupyterHub deployment. The blue boxes denote software in the Pangeo stack that is maintained and developed primarily by the Pangeo community outside of NCAR but to which Xdev contributes, as well. The orange boxes represent software and services that our JupyterHub deployment uses that are hosted by the HPC Division within CISL. Since these services and software require system administrator privileges, Xdev cannot contribute to those technologies without collaboration with HPCD. The red box indicates software and services provided by the Information System’s Division in CISL, another strong collaboration with Xdev, and the green boxes represent software and technology developed by other teams within the Technology Development Division in CISL.


Fig. 4 A diagramatic view of the Pangeo stack and how Xdev can collaborate with other elements within CISL around these software elements

We envision Xdev’s primary role within this ecosystem of software and services as being the development and testing team tasked with improving the user experience on the JupyterHub system through new features and capabilities. Xdev will augment the capabilities of HPCD and ISD, as well as collaborate with other elements of TDD, such as GeoCAT, to achieve these goals.

Core Contributions to Pangeo Technologies

A great deal of our mission is dependent on core Pangeo technologies like Xarray, Dask, and Jupyter. Building third-party packages that have Xarray or Dask as dependencies requires constantly making calls about whether certain features should be implemented in the third-party package itself or in the upstream dependency. Since Xarray and Dask are open source, adding features directly to the upstream dependency is not only possible but encouraged. This mode of development helps keep the core technologies, like Xarray and Dask, healthy and as broadly applicable as possible. Thus, it helps the development of our own technology, which uses Xarray and Dask as dependencies, to contribute to the core technologies ourselves. Obviously, such contributions should also be matched with contributions to the core technology documentation and training materials, as necessary.

The use of these dependencies in our own software also requires that Xdev stay aware of changes in our software’s dependencies (e.g., Xarray and Dask) so that we can continue to ensure our own software’s functionality. This obviously means staying aware of functionality changes and improvements in new releases, but it also means staying aware of issues and bugs that might require fixes in order to address problems propagated through the dependency into our own software. In an open source community, we may be the best people to implement those core technological fixes, and we should be ready to do so, if necessary.