SQR-032: Rendering and testing examples and tutorials in LSST documentation

  • Jonathan Sick

Latest Revision: 2019-08-26

1   Abstract

Examples and tutorials are important elements of successful documentation. They show the reader how something can be accomplished, which is often more powerful and effective than a mere description. Such examples are only useful if they are correct, though. Automated testing infrastructure is the best way to ensure that examples are correct when they are committed, and remain correct as a project evolves. Recently, in DMTN-085 [1], the QA Strategy Working Group (QAWG) issued specific recommendations to improve how examples are managed and tested in LSST’s documentation. This technote analyzes these recommendations and translates them into technical requirements. Subsequently, this technote also provides an overview of how example code management and testing has been implemented.

2   QAWG Recommendations relevant to examples in documentation

The work planned and described in this technote was prompted by recommendations made by the QAWG in DMTN-085. In short, the QAWG recommended that LSST DM adopt a systematic approach to producing, maintaining, and testing examples in our end-user documentation projects. For reference, the relevant recommendations are:

QAWG-REC-13
Provide a central location where examples, scripts and utilities which are not fundamental to pipeline execution are indexed and made discoverable.
QAWG-REC-14
The Project should adopt a documented (in the Developer Guide) policy on the maintenance of example code.
QAWG-REC-15
The Project should prioritize the development of a documentation system which makes it convenient to include code examples and that tests those examples as part of a documentation build.

Of these recommendations, the bulk of the work is associated with QAWG-REC-15. It is through QAWG-REC-15 that we are mandated to specify the conventions and formats for committing examples and scripts into the LSST codebase, and provide the infrastructure for displaying these examples and scripts in end-user documentation. It is also through QAWG-REC-15 that we are mandated to provide infrastructure to test examples and scripts for correctness and reproducibility. We interpret QAWG-REC-13 as requesting that examples and scripts be conveniently displayed and listed through end-user documentation, to solve the issue that examples and scripts are generally not discoverable of both our own developers and the user community. QAWG-REC-13 can be accomplished in conjunction with QAWG-REC-15.

Finally, QAWG-REC-14 requests that we establish a policy for how example code is maintained. The issue is that examples can be broken when developers add or change APIs in the codebase and it is not necessarily the responsibility of those authors to also update all examples and tutorials elsewhere in the LSST documentation and codebase. The work done to fulfill QAWG-REC-13 and QAWG-REC-15, by creating a systematic infrastructure for handling examples and scripts, can assist with this process by alerting developers when their code changes are affecting documentation. Beyond that, in QAWG-REC-14 we must still provide a system for organizing the work associated with breakages in example code.

3   Review of LSST’s existing ad hoc scripts

Although the main focus of this work is associated with examples and tutorials, QAWG-REC-13 also specifically requests that LSST’s ad hoc scripts be systematically included in the documentation. This section summarizes the state of ad hoc scripts in the LSST codebase.

What are these ad hoc scripts?

First, ad hoc scripts are not command-line tasks (soon to be overhauled with the Generation 3 middleware project), which are included in the bin.src directories of packages and that use the lsst.pipe.base.CmdLineTask infrastructure. These scripts are already documented using the Task topic type in LSST documentation. Such task scripts are consistently documented, and centrally indexed. [1]

[1]Tasks are listed in package documentation homepages and there are plans, described in DMTN-30 to list task by theme so that users can discover appropriate tasks without having to be familiar with the structure of the LSST codebase.

3.1   Python scripts in bin.src

Other scripts, outside the task framework, are also included in the bin.src directories of packages. We put Python scripts in bin.src because they are processed by SCons at build time and installed into the user’s $PATH so that they can be executed directly on the command line. For example, the verify package has three scripts in its bin.src directory: dispatch_verify.py, inspect_job.py, and lint_metrics.py. Some of these scripts use the standard argparse package to process command-line arguments (dispatch_verify.py and lint_metrics.py), while others directly parse command line arguments in an ad hoc manner (inspect_job.py).

3.2   General-purpose scripts in examples/

Some useful utility scripts are distributed in the examples directory of packages, rather than bin.src. For example, the plotSkyMap.py script, found in the examples directory of the skymap package uses argparse and is a general purpose utility that is potentially testable and usable. The problem with scripts like plotSkyMap.py is that they are not installed for users. Instead, users need to reference them by their absolute path, often using environment variables created by EUPS:

python $SKYMAP_DIR/examples/plotSkyMap.py

It makes sense to move all command-line scripts into the bin.src directory so that they can be addressed by users without having to know about their association with EUPS packages:

plotSkyMap.py

3.3   Non-reusable utility scripts in examples/

Another category of scripts in examples is non-reusable utility scripts. For example, the pipe_analysis package includes a script called parseLogs.py in its examples directory. parseLogs.py isn’t intended to be used directly since it is hard-coded in a way that is very specific at not reproducible:

logRootDir = "/tigress/HSC/users/lauren/"
lsstTicket = "DM-6816"
hscTicket = "HSC-1382"
rcFields = ["cosmos", "wide"]
bands = ["g", "r", "i", "z", "y", "n921"]
allBands = "HSC-G^HSC-R^HSC-I^HSC-Z^HSC-Y^NB0921"
bandStrDict = {"g": "HSC-G", "r": "HSC-R", "i": "HSC-I", "z": "HSC-Z", "y": "HSC-Y", "n921": "NB0921", "HSC-G^HSC-R^HSC-I^HSC-Z^HSC-Y^NB0921": "GRIZY9", "HSC-G^HSC-R^HSC-I^HSC-Z^HSC-Y": "GRIZY"}
colorsList = ["gri", "riz", "izy", "z9y"]

That said, parseLogs.py is clearly meant to be an executable utility script rather than an example for documentation because it solves a specific problem and doesn’t seem to have a broader didactic purpose.

Example scripts such as this one pose a problem for fulfilling QAWG-REC-15 because that code can only be executed in a single, non-reproducible environment. Such code needs to be re-engineered, including providing a proper command-line interface, if we can hope to use and test it.

3.4   Examples as scripts in examples/

We also see many executable scripts in examples directories that are associated with documentation. An example is the runRepair.py script in the pipe_tasks package. That script is associated with a page in the Doxygen documentation. The reason it’s a script is to be runnable: the script sets up a mock dataset and then runs the RepairTask task on it. This type of script fits the purpose of the original examples framework, but there is a clear mandate from QAWG-REC-15 to improve how these examples are created, managed, and tested. Didactic scripts will be considered later in this technote as part of the examples portion of the work.

3.5   Scripts in languages other than Python

Not all ad hoc scripts are written in Python. For example, run_ci_dataset.sh in the bin directory of ap_verify is a shell script that provides a higher-level interface to the ap_verify.py script.

4   Plan for consolidating and documenting scripts

Based on the review of existing ad hoc scripts in the LSST codebase, we can fulfill QAWG-REC-13 (in relation to scripts) by introducing a systematic approach to including and documenting scripts in the LSST codebase.

4.1   Action: move all utility scripts to bin.src/ or bin/

The first improvement we can realize is by ensuring that any executable script provided with an LSST package is shipped in its bin.src or bin directory. [2]

[2]

This plan of action applies to EUPS packages. LSST code that is packaged for PyPI with setuptools should instead use the console_scripts entrypoints feature to install scripts from a package’s modules:

setup(
    # ...
    entry_points={
        'console_scripts': [
            'cliname = package.module:main_function',
        ]
    }
)

This has the same effect as putting scripts in the bin.src directory of an LSST EUPS package.

This will have the effect of making it possible for users to execute scripts without having to address EUPS environment variables. Using the bin.src directory specifically for Python-based scripts has the benefit of ensuring that the hash-bang is re-written properly to account for SIP security features in macOS. Non-Python scripts can go directly in the bin directory because shebangtron only updates the hash-bangs of Python scripts.

4.2   Action: mandate that the core code from scripts should reside in the main package for testability

Rather than putting the entirety of a script’s code in the script module itself, which resides in bin.src, we should encourage developers to put the entirely of a script’s code in the main package. A useful pattern might be to create a standardized subpackage called scripts [3] within any Stack package. Then inside the scripts subpackage, each “script” is a module that’s named after the command line executable in the bin.src directory. Then the script imports that main entrypoint:

#!/usr/bin/env python
from lsst.some.package.scripts.a import main

if __name__ == '__main__':
    main()

The main function is then responsible for parsing command-line arguments and running the business logic, though ideally main itself is factored such that the core logic is performed in functions that are independent of the command-line context. With this architecture, the script’s internal logic can be tested using the existing unit testing infrastructure (unittest tests run by pytest though SCons). This architecture is already effectively used by command-line tasks. Their executable command-line scripts look like this:

 #!/usr/bin/env python
from lsst.pipe.tasks.processCcd import ProcessCcdTask

ProcessCcdTask.parseAndRun()

The interaction of command-line arguments with script flow can even be tested within unittest-based tests by mocking the output of argparse.ArgumentParser.parse_args. Interactions with other types of external resources can also be mocked.

In summary, by restructuring scripts we can provide comprehensive unit tests for those scripts without having to treat scripts as a special case for testing.

[3]An alternative choice to scripts would be bin. The lsst.verify package has a subpackage called bin.

4.3   Action: scripts are documented in package documentation

Similar to how every function has a page in the Python API reference, and every task has a corresponding Task topic page, every script or command-line executable must have a corresponding documentation page. The structure of this page should be designed and templated equivalently to a topic type. These documentation pages should be listed both on the package’s homepage, and in a central index accessible from the https://pipelines.lsst.io homepage (to be specific) that gathers executables from all packages. The script topic will use Sphinx extensions to auto-populate documentation from the script’s code (see the next action).

The script topic should also provide a way to add metadata about a script, such as a description or tags, to facilitate a useful index of scripts.

4.4   Action: adopt argparse for command-line scripts to enable auto-documentation

To facilitate automatic documentation of command-line interfaces, scripts should use standard frameworks for establishing their interface rather than directly parsing sys.argv. For example, the sphinx-argparse Sphinx extension can automatically document a command-line interface based on the argparse.ArgumentParser configuration for a script. Since argparse is already adopted by the DM Style Guide, this recommendation should be non-controversial. Nevertheless, some simple scripts have been written to avoid argparse, and those scripts should be ported to argparse to facilitate documentation.

To work with sphinx-argparse, scripts need to be written such that the argparse.ArgumentParser is returned by a function that takes no arguments:

def main():
    parser = parse_args()
    # ...


def parse_args():
    parser = argparse.ArgumentParser(description='Documentation for the script')
    parser.add_argument('--verbose', default=False, help='Enable verbose output)
    return parser

Such a requirement will need to be added to the DM Python Style Guide.

5   Survey of examples and tutorials in LSST documentation

In the second part of this technote, we consider example and tutorial code that appears in documentation, and attempt to provide a road map for fulfilling QAWG-REC-13 and QAWG-REC-15. As with the first part, concerning utility scripts, we first review the current landscape of example code, and in later sections identify technologies and actionable steps towards meeting QAWG-REC-13 and QAWG-REC-14.

5.1   Examples in examples/ directories

The examples directory does, in fact, contain example code (though see also 3.2   General-purpose scripts in examples/). Examples exist in many forms: C++ source and header files (.cc and .h), Python modules, and Jupyter notebooks.

In the most successful cases, the Python and C++ files are referenced from documentation text. Originally, LSST Science Pipelines documentation was written in Doxygen and the .dox files and docstrings included the contents of files from examples. For example, the measAlgTasks.py example is referenced from the docstring of the SourceDetectionTask. Newer documentation being written in reStructuredText is merely linking to the GitHub blob URLs of files in examples because the multi-package build prevents examples from being available at a fixed relative URL during the build process.[4]

[4]See the Overview of the pipelines.lsst.io build system in Documenteer’s documentation.

Many of the Python examples, and generally all of the C++ examples, are structured as executables. In the case of the Python examples, the command-line interface provides a toggle for activating the debug framework or to optionally open a display (see measAlgTasks.py). Thus these examples are distinct from ad hoc scripts that are also found in the examples directory.

Not all examples are referenced from the documentation, though. For example, the statisticsMaskedImage.py module in the examples directory of afw is not referenced in any documentation, despite seeming to be genuine example.

5.2   Test programs in examples/ directories

In addition to examples that are associated with documentation, some files in examples directories (typically C++) are neither true examples nor ad hoc utilities. These files seem to be tests of an ad hoc nature. Examples of this are the ticket647.cc and maskIo2.cc programs in afw. The former appears to reference a ticket from the deprecated Trac system, and maskIo2.cc seems to test memory management in C++ code.

DM does have a testing system for C++ code. Ad hoc tests, like the ones identified, above should be moved out of examples and into the tests directory.

5.3   Data files in examples/ directories

In rare cases, data files are located in the examples directories of packages. One such file is NewSuprimeCam.paf in afw, which has no references anywhere in the afw codebase. Orphaned files such as this one should be removed.

5.4   Python doctests

Another category of example code that is commonly found in Python are doctests, which are supported by Python itself through the doctest standard library package. doctests are formatted like interactive Python sessions, and show both the input that a user might enter at a Python prompt, along with the expected output. For example:

>>> a = [1, 2, 3]
>>> a.append(4)
>>> a
[1, 2, 3, 4]

doctests can be found in docstrings (particularly the Examples section of a Numpydoc-formatted Python docstring), as well as in reStructuredText files. Because docstrings show both inputs and outputs, they work well as testable examples because test harnesses can run the example and verify that the output matches the expected output.

5.5   Untested Python and shell samples in reStructuredText

Besides doctests, code samples can also be added to documentation with reStructuredText directives such as code-block, literalinclude, and prompt. For example, the Getting Started tutorial series in the Pipelines documentation uses code-block directives to include both shell commands and their output, along with Python scripts.

5.6   Jupyter notebooks

Jupyter Notebooks, like doctests, are well suited for creating testable documentation because of how they mix prose, code, and output cells. Jupyter Notebooks are particularly notable for the immediacy and interactivity of their writing process.

LSST uses Jupyter Notebooks in a number of contexts, including as documentation. As mentioned before, Jupyter Notebooks appear in the examples directories of packages. Entire Git repositories are also dedicated to collecting Jupyter Notebooks. For example, the notebook-demo repository contains demo notebooks for LSST’s Nublado platform. At the moment, Notebooks aren’t part of Sphinx documentation builds.

6   Consolidation of approaches to examples

In the previous section, we reviewed the various types of examples that exist in LSST codebases. Given the number of formats that examples can currently be found in, it’s beneficial to consolidate our usage to a defined set of technologies and methodologies that are both convenient to integrate into documentation (addressing QAWG-REC-13) and test (addressing QAWG-REC-15). The QAWG recommended that one technology should be adopted:

There are various technologies which could be adopted to address this goal.[5] The WG suggests that standardizing upon a single technology is essential, but takes no position as to which technology is most appropriate.

[5]For example, Jupyter notebooks or Sphinx doctests.

Although adopting a single technology is appealing, such a limitation may prove inappropriate for the types of documentation that LSST writes, and the types of things that are documented. The approach that we will pursue in this technote is to address the specific scenarios where examples are written for LSST documentation, and to associate a specific approach with that scenario. This way, even though we support multiple technologies, only one is permitted for each documentation scenario. We believe that most scenarios of writing examples in documentation can be covered with two technologies: Python doctests and Jupyter Notebooks. C++ examples remain a special case, and will be supported by a third approach. The following sections review these adopted technologies and the scenarios that they support.

6.1   Doctests

Doctests are standard in Python, have have robust support in both Sphinx (the tool that generates our documentation websites) and in pytest (the tool that runs our Python unit tests). The Astropy project is an excellent example of doctest-driven documentation. By using doctests, the Astropy project provides ample examples of their APIs, and these examples are tested automatically as part of their continuous integration process.

Doctests excel in their integration with existing software and documentation development practices. Doctests are convenient to add to docstrings of Python functions, classes, and methods. Doctests are also convenient to add to reStructuredText files, which is where the bulk of LSST’s user and conceptual documentation is already written. For example, each task already has a task topic page written in reStructuredText. Doctests demonstrating that task as a Python API can be added directory into that reStructuredText file. Being plain text, doctests are convenient to review as part of a pull request.

Compared to Jupyter Notebooks, doctests are slightly less convenient to write. Instead of the writing and execution environment being combined, developers may choose to write Python statements in a scratch Jupyter Notebook or IPython shell and copy the source and output into a doctest. Testing the doctest also requires running a command. However, given the success and abundant use of doctests in projects ranging from the Python documentation to Astropy, it would appear that workflow issues are not significant.

We recommend that doctests be adopted for docstrings and for how-to documentation written in reStructuredText where it is important for the example to integrate seamlessly with the existing documentation.

6.2   Jupyter Noteboooks

Jupyter Notebooks are the second technology that would be good to consolidate towards. In many ways, Jupyter Notebooks have similar attributes to doctests in that they combine prose, source code, and outputs. Compared to doctests, notebooks add a few additional capabilities: support for running shell commands, and integration as a development and execution environment for both writers and readers. Given the adoption of Jupyter notebooks by the LSST Science Platform, it’s also obvious that notebooks cannot be ignored as a platform for creating examples.

Notebooks do have some disadvantages compared to doctests that prevent us from adopting them as the dominant technology for all examples. First, their JSON format is difficult to integrate into Pull Request workflows where merge conflicts can be expected. Similarly, notebooks require a working Jupyter server to view and edit, as opposed to the minimalist requirements of doctests. As the LSST Science Platform becomes more mature, it will become easier to include Jupyter Notebooks in a development workflow.

Second, notebooks are also difficult to integrate into Sphinx documentation. Markdown is the primary prose format for Jupyter Notebooks. Although it’s possible to write in reStructuredText, it won’t be rendered in the browser. This means that notebooks cannot take advantage of Sphinx’s cross-referencing syntax. Nor can reStructuredText files use cross-reference syntax to link to a Jupyter Notebook. For this reason we suggest that it’s better to not deeply integrate Jupyter Notebooks within a Sphinx documentation page. Instead, Jupyter Notebooks ought to be standalone pages.

In other words, Jupyter Notebooks work well for delivering tutorials. In What nobody tells you about documentation, the author Daniele Procida describes four distinct types of documentation:

Reference
A comprehensive description of the product.
Explanations
Content that helps build understanding, background, and context.
How-to guides
Specific recipes, often with steps, that describe how to accomplish a common task.
Tutorials
A learning-oriented lesson.

Doctests work well when integrated with reference documentation (examples in docstrings, for example), in how-to guides, and to a lesser extent in explanatory guides as well. That type of documentation is deeply integrated with reStructuredText and Sphinx.

Jupyter notebooks on the other hand are excellent for tutorials because, as a lesson, they can stand apart from the main body of the documentation. Sphinx’s features, such as custom syntax, are not generally needed for tutorials. The reader’s ability to download the notebook itself and follow along and make experimental adjustments to the tutorial is hugely beneficial. Lastly, tutorials experience less churn during regular development than other types of documentation, which makes the notebook’s requirement that it cannot be edited in a text editor more acceptable. Thus, we recommend Jupyter notebooks as an ideal medium for producing tutorials.

Jupyter notebooks are also useful for other types of documentation that benefits from an integration with software. For example, technical notes could be written as Jupyter notebooks. The nbreport platform is also built around the concept of using Jupyter notebooks to generate regular reports augmented with templated computations.

6.3   C++ examples

Together, doctests and Jupyter Notebooks cover scenarios for most of the examples that LSST might want to write. One scenario that isn’t addressed, though, are C++ examples that are currently found in the examples directories of packages. Neither doctests nor Jupyter Notebooks support C++ code.

For C++ examples, it may best to continue the existing pattern of placing source files in the examples directory, having SCons run the compilation of those examples, and referencing those examples from the documentation. Note that displaying files from the examples directory still needs to be accommodated in the Sphinx builds as mentioned in 5.1   Examples in examples/ directories.

6.4   Summary of documentation scenarios and technologies

In summary, these are technologies that DM should adopt to produce examples and the appropriate scenarios for each technology:

Python doctests
  • “Examples” sections of Python docstrings.
  • Python API demos and how-tos integrated with reStructuredText/Sphinx documentation.
  • Pure-Python tutorials written in reStructuredText/Sphinx.
Jupyter Notebooks
  • Standalone tutorials that use Python and/or the shell, which are associated with a Sphinx documentation site.
  • nbreport [2] templates and instances.
  • Technical notes written entirely as a Jupyter Notebook.
Files in examples
  • Examples written in C++ that are referenced from reStructuredText/Sphinx documentation.

6.5   Types of examples not directly covered by adopted technologies

Some scenarios are not well covered by the adopted technologies. These are:

  • UI-based tutorials
  • Project-building tutorials

LSST will use UI-based tutorials in documentation of the Science Platform. There isn’t a technology that combines the content of a UI-based tutorial with a machine-testable plan. In industry, UI-based tutorials are often periodically reviewed and updated by a QA or documentation team. It’s conceivable that UI tutorials could be co-developed with a Selenium test script (or similar). Selenium is also commonly used in industry to generate screenshots for UI-based tutorials since it’s often the appearance of the UI that changes most frequently.

Project-building tutorials are a common format for developer tutorials where the reader is guided through building a project consisting of multiple source files that are incrementally built up. tut is a Sphinx extension that provides an approach to creating a project-based tutorial in Sphinx. It works with a Git repository where each branch contains the code at each stage of the project.

7   Approaches for integrating doctests with Stack testing

Doctests are one of our adopted technologies for writing testable examples. This section considers the various approaches for running and validating doctests as part of either the general software testing process or the documentation build. In general, there are two types of approaches: run doctests through pytest while the software is being tested, or run doctests through Sphinx when the documentation is built.

7.1   Running doctests through pytest

The main testing command for the Stack (which the LSST Science Pipelines is part of) is scons tests. SCons, in turn, runs pytest, which provides test discovery, execution, and reporting. Integrating doctests with pytest, and thus scons tests is appealing because it would enable us to test doctests without changing developer workflows.

Pytest supports doctests through a --doctest-modules command-line argument. In principle, pytest should discover all Python modules in the package and run their doctests, similarly to how it discovers test modules and executes them.

As a case study, the verify package uses doctests that are run by pytest using its --doctest-modules argument. Note that in order for modules to be discovered, we had to specify the python, tests, and bin.src directories where modules can be found in a standard LSST EUPS package. In the setup.cfg file, pytest is configured as:

[tool:pytest]
    addopts = --flake8 --doctest-modules python bin.src tests
    flake8-ignore = E133 E226 E228 N802 N803 N806 N812 N815 N816 W504

The disadvantage of this approach is that the specification of python bin.src tests as default options through the setup.cfg file prevents a developer from easily running pytest against a single test module. Additional work is needed to understand why pytest cannot automatically discover LSST’s Python modules by default.

In addition to Python modules, pytest can also gather and run doctests in reStructuredText files using the --doctest-glob argument. For example: --doctest-glob=doc/**/*.rst would test all reStructuredText files in a package’s documentation.

7.2   Enhancing pytest with Astropy’s pytest-doctestplus

Astropy created a extension for pytest called pytest-doctestplus that enhances pytest-based doctest testing. Its main features are:

  • Processing doctests in reStructuredText files (which is now handled natively by Pytest).
  • Approximate floating point comparison.
  • Advanced doctest skipping control for modules.
  • Integration with the pytest-remotedata plugin to enable skipping tests that require remote connections.

The floating point comparison feature is useful for avoiding test failures because of small floating point rounding differences between a doctest and an execution. It handles cases like this:

>>> 1. / 3.  # doctest: +FLOAT_CMP
0.333333333333333311

Such a directive is likely useful to enable by default.

pytest-doctestplus allows developers to indicate that any doctests associated with a Python class, function, method, or whole module should be skipped through a __doctest_skip__ module-level variable.

For example, to skip all doctests in the function get_http in a module:

__doctest_skip__ = ['get_http']

It also supports wildcard matching of names:

__doctest_skip__ = ['HttpClient.http_*']

An entire module can be skipped with a module-level wildcard:

__doctest_skip__ = ['*']

pytest-doctestplus provides a similar module-level variable to configure doctests that should be skipped if an optional dependency is not present.

In the context of reStructuredText pages, and within docstrings, pytest-doctestplus also provides several directives [6] for skipping tests and creating invisible doctests for environmental set up.

[6]These reStructuredText directives are not actually part of pytest-doctestplus. In fact, they’re shipped as part of the sphinx-astropy project as sphinx_astropy.ext.doctest. This means that adding pytest-doctestplus also requires adding sphinx-astropy as a dependency.

The testsetup directive runs the code within the doctest, but does not display the code in the published documentation:

.. testsetup::

   import lsst.afw.table as afwTable

testsetup, as the name implies, is useful for running setup code that would distract from the narrative of the page itself.

The doctest-skip directive renders the doctest, but skips executing and therefore testing, the code:

.. doctest-skip::

   >>> 1 / 0

Finally, there is also a special comment that ignores all doctests on a page:

.. doctest-skip-all

Overall, pytest-doctestplus appears to be a useful extension of pytest’s basic doctest capability, and should be part of our solution for testing doctests.

7.3   Sybil

An alternative to pytest’s --doctest-modules mode and pytest-doctestplus is Sybil. Sybil provides additional features for testing Python examples in reStructuredText/Sphinx documentation.

7.3.1   Features

The main use case for Sybil over other systems is being able to build custom example parsers. Whereas pytest and pytest-doctestplus only check for traditional doctests, Sybil provides additional parsers to check examples written in other types of syntax, such as code-block directives.

This flexibility is useful in cases where an author might write a function or class in a code-block directive, and then use a doctest to exercise that example class or function. The code from both the code-block and doctest are treated as part of the same namespace.

In addition to the code-block parser, Sybil provides an API for additional parsers should we wish to provide examples in custom reStructuredText directives or in different languages or syntaxes. For example, Sybil could operate on bash scripts or commands:

.. code-block:: sh

   $ echo Hello world
   Hello world

Similarly, Sybil could also be extended to validate YAML or JSON-format code blocks.

Sybil also provides an invisible-code-block reStructuredText directive:

.. invisible-code-block:: python

   import lsst.afw.table as afwTable

This directive can be used to execute code within the namespace of the page’s tests without rendering content onto the page. Used judiciously, invisible-code-block could be useful for adding setup code and environment assertions to ensure that the examples are testable without adding distractions for readers.

In addition, Sybil provides a flexible skipping mechanism. Using a skip reStructuredText comment, single examples or ranges of examples can be skipped.

.. skip: next

>>> 1 / 0

A range of examples can also be skipped:

.. skip: start

>>> 1+1
2

Some text...

>>> 40+2
42

.. skip: end

Examples can also be skipped based on a logical test (the invisible-code-block directives provide a place to write auxiliary code for these tests).

.. invisible-code-block::

   import os
   is_travis = os.getenv('TRAVIS') == 'true'

.. skip: is_travis is False

>>> os.getenv('TRAVIS')
'true'

7.3.2   Integration with pytest

Sybil integrates with pytest, among other Python test runners. To use Sybil with pytest, we would add a conftest.py file to the doc directories of packages (or any other documentation project). In this conftest.py file we can configure features like the parsers mentioned mentioned in the previous section.

By executing pytest from a package’s root directory, as SCons already does, pytest should automatically discover the doc/conftest.py file and begin testing the doctests in the reStructuredText source. Thus Sybil can integrate well into DM’s existing pytest-based testing system.

Finally, as alluded to above, Sybil is pitched squarely at running doctests in reStructuredText files, not in docstrings within Python modules. Thus Sybil would need to be used in conjunction with pytest itself or pytest-doctestplus to test docstrings.

7.4   Testing doctests with sphinx.ext.doctest

Another method of testing doctests in documentation is as part of the Sphinx documentation build, rather than as part of the unit testing with pytest. sphinx.ext.doctest, a Sphinx extension included with Sphinx, provides this capability.

sphinx.ext.doctest provides three methods for writing doctests:

  1. Regular doctests that use the >>> syntax.
  2. A doctest directive that provides control over test groupings, what doctest directives are applied to process the doctest, and whether or not to hide the doctest in the build site.
  3. A testcode and testoutput directive pairing that allow writers to separate the block that displays the example code from the block that displays output.

This third method is unique to sphinx.ext.doctest. It gives authors the flexibility to separately introduce the input and output. On the other hand, readers are used to seeing code and output paired together (not only do doctests pair code and output, but Jupyter Notebooks as well).

In addition to directives for writing the examples themselves, sphinx.ext.doctest also provides testsetup and testcleanup directives. The content of these directives is automatically hidden, and are automatically run before and after, respectively, the test groups that they are associated with. Similar to the invisible-code-block directive provided by Sybil, the testsetup directive can both run preparatory code and also add variables to the namespace that can be used by the examples.

Finally, sphinx.ext.doctest provides means for conditionally skipping tests of examples.

7.5   Summary of doctest testing approaches

There are several approaches to validating doctest-based examples that differ based on how they are run, where they run, and what features they add beyond the basic doctest library.

Of the options listed, we immediately dismiss the strictly pytest-driven approach as pytest-doctestplus provides an equivalent, but improved, feature set. We also dismiss the sphinx.ext.doctest approach because it occurs during the documentation build, rather than during regular testing (driven by pytest). We believe that documentation testing with pytest-based testing brings documentation integrity to the forefront of developer culture, and helps ensure that breakages of examples are more visible. This reasoning also weights heavily in our chosen approach to testing Jupyter Notebooks.

The two approaches that are worthwhile to pursue are pytest-doctestplus and Sybil. The pytest-doctestplus extension is advantageous because it can be used for both doctest examples in docstrings and in reStructuredText pages. As such, pytest-doctestplus provides a uniform system for testing all doctests that developers write.

On the other hand, Sybil only operates on reStructuredText pages. Compared to pytest-doctestplus, though, Sybil has an extensible parsing system that can allow us to test examples in other languages, such as shell scripts or even HTTP calls. Sybil also has a more flexible syntax for constructing invisible code blocks and for skipping tests.

There are fundamentally two options for applying doctests, then:

  1. Use pytest-doctestplus to test all doctests.
  2. Use pytest-doctestplus to only test doctests in docstrings, and then use Sybil to test reStructuredText pages.

From a capability standpoint, the second option is more appealing as it allows greater flexibility in testing examples in reStructuredText documentation. The drawback is that developers must be aware of the distinction of writing doctests in these two environments, and learn the skip syntax for each.

8   Testing Jupyter Notebooks

8.1   nbval for testing notebooks with pytest

nbval enables pytest to run on Jupyter Notebooks. It determines whether the Jupyter Notebook, when re-executed, can reproduce the outputs stored in the notebook. In this way, nbval treats Jupyter Notebooks as sophisticated doctests.

nbval provides different levels of control over how the output stored in the Jupyter Notebook is compared against output from executing the notebook in a test environment:

  1. All cells can be tested by running as pytest --nbval.
  2. Only cells specially marked with a # NBVAL_CHECK_OUTPUT marker comment can be tested by running as pytest --nbval-lax.
  3. Checking all cells, but only after “sanitizing” the reproduced and stored outputs to avoid testing outputs that are intrinsically semi-random.

The --nbval-lax mode is a low buy-in means of testing notebooks by allowing authors to mark just those cells that are representative and known to be testable. Note that their is a companion comment, # NBVAL_IGNORE_OUTPUT that causes nbval to skip testing a cell. This is a useful escape valve for cells that are difficult or impossible to test.

The sanitization approach is more technically involved. In this mode, we would provide an ini-format file with regular expressions and replacement strings. nbval runs these regular expressions over the outputs and replaces the matched strings with a simplified replacement string that is tested against.

nbval has additional advanced features that are useful. One is the # NBVAL_RAISES_EXCEPTION code comment that permits a cell to raise an exception, and directs nbval to test the traceback.

Instead of using Python code comments (which are user-visible), cells can also be annotated with Jupyter-native tags.

Finally, nbval can be configured to skip certain output types, such as stderr or application/javascript.

nbval is known to be compatible with the xdist plugin for running tests in parallel. In this case, an entire notebook would be run together as a unit.

It is not currently known how to control which Jupyter kernel pytest or nbval runs notebooks with, or whether this is configurable.

Overall, nbval is an excellent and comprehensive solution for ensuring that Jupyter Notebooks are reproducible. One caveat that must be recognized, though, is that nbval requires that notebooks be committed into documentation repositories with their outputs committed. This pattern runs counter to the practice of always stripping notebooks of outputs before committing them to a Git repository. Committing outputs increases the probability of merge conflicts should multiple authors be working on the same notebook simultaneously in separate branches. This reinforces the notion that Jupyter Notebooks should only be used in special circumstances, such as tutorials, rather than as the primary format for all of LSST’s documentation and examples.

8.2   nbpages and nbconvert as a testing device

The one potential drawback of the nbval/pytest approach, described above, is that notebooks must be committed with their outputs. As discussed, committing outputs makes Git diffs more difficult to interpret and increases the difficulty associated with resolving merge conflicts. nbpages, developed at the Space Telescope Science Institute, attempts to work around this issue by creating a notebook publishing workflow that rigorously uses notebooks committed to Git without outputs.

Essentially, nbpages is a front-end to nbconvert. nbpages runs nbconvert on each notebook in a repository, which executes the notebook programatically, and then converts the notebook into an output format such as HTML or reStructuredText. As such, nbpages provides smoke-test level testing of notebooks.

Obviously, without having existing outputs, it is impossible for this method to discern whether the outputs are correct or not. However, simply running the notebook programatically protects against notebooks that outright do not run.

8.3   Summary of notebook testing approaches

Overall, nbval is a comprehensive solution for testing Jupyter Notebooks intended for documentation, and integrates into our existing pytest workflow. An nbconvert-based testing approach is not as compelling as actually validating the reproducibility of a notebook’s outputs. Further, the fact that nbval relies on notebooks where the outputs are included can be seen as a feature since it means that a documentation site can be regenerated without having to re-run the notebook itself. This is a useful capability so that developer builds can run without a massive computational investment. In summary, we should adopt nbval for projects such as the pipelines.lsst.io and nb.lsst.io documentation projects.

9   Execution environments for testing examples and scripts

Since examples (whether they are are doctests or notebooks) and scripts are tested through pytest, the environment where they are tested corresponds to the environment where the corresponding software is tested or documentation is built. This section considers the specific execution for different LSST DM projects.

9.1   The LSST Science Pipelines test environment

The LSST Science Pipelines is tested by ci.lsst.codes, a Jenkins CI deployment. Developers run the stack-os-matrix job to test the LSST Science Pipelines during their regular development. Since stack-os-matrix runs the scons command, which in turn runs pytest, tests for doctests, notebooks, and scripts will be executed in the stack-os-matrix environment.

Parallelism is provided “for free” by virtue of the pytest xdist plugin.

Any datasets that are referenced by the examples must be available in the stack-os-matrix job as well.

9.2   The Nublado test environment

Examples for the LSST Science Platform generally assume access to the capabilities of the LSP. For example, an example notebook for the notebook aspect will assume access to mounted filesystems and data access APIs. These example notebooks will also assume that certain software is installed, whether they are command-line tools or Python packages. This necessitates that examples for the LSP should be tested on the LSP itself. Mocking the LSP is not practical.

To do this, we envision adding a capability to Nublado (the LSST JupyterLab-based platform) that allows API-based execution. That is, through an HTTP API, the documentation build platform should be able to request that Nublado start up a headless pod running JupyterLab and execute a command. Such a command might do this:

  1. Clone a Git repository (at a specific commit) that contains the documentation being tested.
  2. Temporarily install additional Python dependencies needed for the documentation and its tests.
  3. Run a shell command (likely pytest) to execute the tests.
  4. Gather the output of the tests and return it to the caller.

In this scenario, the documentation would be built in a CI environment (likely ci.lsst.codes), but the tests would run in Nublado itself. Another interesting possibility is that the same API access to Nublado could also be used to build the documentation itself, in addition to testing the documentation. This would be useful for enabling custom Sphinx extensions, that LSST writes, to introspect the Nublado environment and automatically document features such as available Python packages, command-line tools, and datasets on shared file storage.

All of this, however, requires that LSST develop API access to Jupyter to spawn and operate headless JupyterLab pods.

10   References

[1][DMTN-085]. E.C. Bellm, H.-F. Chiang, A. Fausti, K.S. Krughoff, L.A. MacArthur, J.D. Swinbank, and T Roby. The QA Strategy Working Group Report. 2019. URL: https://dmtn-085.lsst.io
[2][SQR-023]. Jonathan Sick. Design of the notebook-based report system. 2018. URL: https://sqr-023.lsst.io