On development models for sharing (experimental) code

One core aim of Sage is to foster code sharing, and to encourage groups of researchers, teachers, and other users to get together to develop new features they need either on top of or within Sage, and share them.

Over the years, many development workflows have been experimented by various groups of people to improve Sage in certain areas, like Sage-Combinat for (algebraic) combinatorics, Sage-Words for combinatorics on words, SageManifolds for differential geometry, purple-sage for number theory, ...

The goal of this document is to discuss the different workflows that have been tried with their pros and cons, to share best practices and to brainstorm about what support and recommendations Sage could provide for various use cases. Eventually, this could become a section of the developer's manual (though this can be of interest for other people wanting to start sharing code without necessarily contributing to Sage), or a page of the sagemath.org website.

At this point this is a collection of notes by N. Thiéry; please hack in and contribute your own vision!

See also:

• A SageMathExternalPackages

• A discussion on sage-devel about modularization of the Sage library

Objectives of a development workflow

Of course the milleage will vary from project to project, but the objectives of a development workflow can typically be to:

1. Support fast paced development within a group of users working on the same topic, or needing similar features.

2. Support rapid dissemination of experimental features.

• The goal is simultaneously to support users, and to get early feedback on the code. Typical needs:
• Using, for a given calculation, experimental features from different areas, developed by different groups of people
• Getting the latest version of a feature, without having to upgrade all of Sage (e.g. just before delivering a talk!!!)
• Feature discovery: increasing the chances for someone to discover that a given feature is being implemented somewhere

3. Foster high quality code by promoting documentation, tests, code reviews.

4. Foster intrinsic high quality code by providing an *ecosystem* where (experimental) code can live, compete with other implementations, mature and be selected or die, all at a fine granularity.

5. Strike a balance between centralized and decentralized.

• In particular mitigate the risks of code-bloat of the Sage library versus the risks of death of code lying out somewhere on the web.

6. Minimize *maintenance* overhead, and in particular code rotting.

7. Remain flexible between the all-in-one versus packages development models (simplifying things out: between Sage's model and GAP's model).

8. Promote extending existing Sage classes and modules with additional features.

• This eases dynamic feature discovery by users (once installed, features can be found at their expected location) and enable transparent migration of code inside the Sage library if and when desired (no need to change the code itself, nor code using it). This also promotes coherent coding standards. Note: subclassing is not always an option to extend a class, e.g. when a feature is to be added to an abstract base class of many concrete classes (subclassing each and every concrete class would be a pain). See also:

• http://www.csd.uwo.ca/~watt/pub/reprints/2006-dsal-postfacto.pdf

• https://en.wikipedia.org/wiki/Extension_method

• https://en.wikipedia.org/wiki/Monkey_patch

Existing workflows

Direct integration into Sage

In this workflow, each feature is shared by integrating it directly into Sage.

Pros:

• Simplicity for the user: all stable features are directly available in Sage
• Simplicity for Sage developers: no additional workflow to learn
• No need to worry about release, distribution, test infrastructure, ...
• Promotes early integration of code and objective 3
• Makes objective 8 straightforward

Cons:

• Limited support for objective 2
• Slows down the development: once a feature is in Sage, any change needs to be reviewed, refactoring of the public API requires taking care of backward compatibility. No good for objective 4
• Getting the latest feature forces updating to the latest version of Sage
• Introduces a bias toward code bloat (in doubt, features tend to be added to Sage)
• When development is faster than reviews, the maintenance effort in having many open tickets gets heavy when minor changes to an early ticket has to be merged into all later ones.

Examples:

• SageManifolds, cf. the metaticket #18528

• ACTIS: Algebraic Coding Theory for Sage, cf. the metaticket #18846

Discussion:

• Soften model using external repo: In the beginning of ACTIS (see above), we maintained a public clone of Sage on Bitbucket where each major feature set was a branch. Once our main design was mature enough, the first few branches were made into Trac tickets and merged in Sage. This fully achieved objective 2 and 4 in this phase. When choosing the scope of a branch, attention was given to minimising dependencies, easing the maintenance burden of parallel development. However, extracting tickets from branches was manual and error-prone, and changes done in the trac review phase were annoying to port back to the public repo. So after the most volatile period of design, we abandoned this model.
• Use the @experimental decorator to mitigate the backward compatibility issue while the code is not yet fully mature. The decorator is a bit clumsy to use due to doc-testing in Sphinx (tricks need to be done to avoid printing the experimental warning on each doc-test), see e.g.

  AsymptoticRing.

Feature branches

In this workflow, features or feature sets are implemented as branches on the Sage sources.

Pros:

• Makes objective 8 straightforward
• Encourages integration into Sage
• Development history is automatically kept upon integration into Sage

Cons:

• Branch needs to be regularly updated to prevent code rotting due to syntactical conflicts with changes in Sage (though automatic merges help)
• Objective 2 requires basic git knowledge from end-users
• Lack of modularity for objective 2: due to potential conflicts, it's not easy to combine features from several branches; upgrading to the latest version of a branch often forces a change of version of Sage
• Cherry picking certain mature features for integration in Sage is somewhat cumbersome (the granularity of branches and commits is orthogonal to the granularity of features)
• It's hard to strike the right granularity in terms of feature / feature set. We tried dependency tracking among branches as a way to build feature sets out of features, but this did not work well
• Because of the above, this workflow does not work well for objective 4
• Introduces a bias toward the all-in-one development model

Examples:

• Global function fields

Features in trac:

Sage development trac can be used to host the features. The ticket for a feature has milestone sage-feature and keeps a feature branch that provides a special functionality that can be merged to the Sage core by the user. The ticket is not reviewed (i.e, not goes into needs review status) until it is accepted as a standard feature and integrated to Sage. The user can select a set of features at build time and pull the feature branches from trac and start to make in the usual way. A special script "make-sage-with features" might be provided to make this easy.

A feature in trac should provide at least

• Description about the feature.
• Info about the author.
• Info about the latest Sage release with which the feature works.
• Info about dependencies, that is, other tickets that this feature depends on.
• Optionally a link to the host at which actual development occurs, like Github repos.

Other remarks:

• The feature branch contain code and documentation.
• The patchbots test the feature against the latest Sage release regularly.
• The features in trac should be either orthogonal or competent to other features in their functionality.
• Some parts of the present Sage library may be turned into features.

Patch queue as used by Sage-Combinat between 2009 and 2013

See also the bottom of this page.

TODO: description

This section is just for reference: there used to be a strong rationale for this workflow with the former Sage development workflow and a given context. But not any more.

Pros:

• Relatively good for objective 1 (except for objective 6)
• Relatively good for objective 2 (thanks to "sage -combinat install"), except for modularity and requiring some Sage recompilation
• Objective 8 is straightforward

Cons:

• Complexity of working at the meta level (version control on the patches)
• Really bad at objective 6: Horrible maintenance overhead due to syntactic conflicts and lack of automatic merging
• Introduces a strong bias toward code death, or at least non integration into Sage
• Monolithic: one could not use several patch queues at once, so this did not support overlaping groups of people working on different topics; this introduced a non-natural barrier between Sage-Combinat and the rest of the world, and prevented rapid reconfiguration of projects around topics and groups of developers

Standalone (pip) packages

Here the idea is to implement feature sets as independent Python packages on top of Sage. Converting a bunch of Python files into such a package to make it `easy to install <http://python-packaging-user-guide.readthedocs.io/en/latest/distributing/>`_ is straightforward.

Examples:

• Template for creating Sage packages

• Modular Abelian Varities

• Python implementation of chebfun

• Purple Sage

• slabbe-0.2.spkg

  See also this blog post Note that this is an spkg, rather than a standard pip-installable package.

NON-Examples:

• SageManifolds

  • The install script for this does all kinds of copying files directly into the sage install, using sed to modify parts of the sage library, etc. It's terrifying. -- William.

    True but that's only a provisory thing for an easy install by a newbie, until the process of full integration of SageManifolds in Sage, started at #18528, is finished. For a developer, the recommended installation process is via git, not via the above script. Actually, we started SageManifolds as a (new-style) spkg and it was distributed as such until version 0.4. Then, in order to ease the review process, we split it in many tickets (listed at #18528) and devise the above script just for end users not familiar with git and make.

    Really the development workflow of SageManifolds pertains to the category Direct integration into Sage above, hence it is a Non-Example here -- Eric. Or maybe to the category "standalone package with an integration mission below". By the way, the usage of scripts could potentially be replaced by the monkey patching approach described below, though I'd need to check the exact use cases. Let's discuss this at some point! -- Nicolas

    UPDATE (15 Jan 2017): SageManifolds is now fully integrated in Sage 7.5. There is therefore no need for an install script, even for a newbie; cf. the new download page -- Eric.

• CHA "It is recommended to use the more recent implementation from the branch attached to this ticket rather than this library."; I think this is just some code to copy into the sage library or run directly, with no package support at all.

Pros:

• Good for objectives 1, 2, 4

Cons:

• Handling of compatibility with various versions of the dependencies (in particular Sage)
• Risk of code rotting (as Sage evolves over time) or death (if it's not maintained)
• Requires coordination with Sage and related packages to not step on each other

Standalone (pip) packages with an integration mission

This is a variant on the previous development workflow, with an explicit focus on easing (or even promoting) the integration of mature code into Sage.

Specifics:

• Layout the code as in the Sage library, with top module called

  e.g. sage-blah instead of sage. For example, to add a method to the Sage class Partition, one would put it in an otherwise empty class sage-blah.combinat.partition.Partition.

• Use recursive monkey patching to insert all the code dynamically in the Sage library, for example by using the

  recursive-monkey-patch pip package The effect is to patch the Sage library, as with branches or patch queues; however this is done semantically at the granularity of methods rather than syntactically at the granularity of lines in the source code.

Examples:

• Sage-semigroups (quite preliminary!!!)

Pros:

• Same as above
• Objective 8 is straightforward
• Lighter maintenance overhead compared to branches or patch queues: one only needs to take care of semantic conflicts, not syntactic ones.
• The integration of mature code into Sage helps for objective 3 and for the maintenance as well: keeping the library as a "small layer" over Sage reduces the risks of irreversibly drifting away, and reduces the amount of updating.
• Depending on how strongly one pushes toward the integration of mature code, one can flexibly interpolate between the all-in-one model and the package model

Cons:

• The concept has not yet been really battlefield tested!
• Moving code into the Sage library is done by copy pasting. This makes for a clean diff showing just the addition of the new methods, but means that one looses the history and author tracking (that's not that different from history squashing as used by many projects)

Standalone (pip) packages that provide native namespace packages

This is another variant, enabled by the transition to Python 3.

See https://trac.sagemath.org/ticket/28925

Specifics:

• Transform packages in the Sage library to native namespace packages

  • by removing init.py

• In the standalone pacakge, lay out the code in a source tree parallel to
  • that of the Sage library.

Examples:

• https://github.com/mkoeppe/sage-numerical-backends-namespace

Historical: what was this Sage-Combinat queue madness about???

Sage-Combinat is a software project whose mission is: "to improve the open source mathematical system Sage as an extensible toolbox for computer exploration in (algebraic) combinatorics, and foster code sharing between researchers in this area".

In practice it's a community of a dozen regular contributors, 20 occasional ones and, maybe, 30 users. They originally collaborated together on a collection of experimental patches (i.e. extensions) on top of Sage. Each one describes a relatively atomic modification which may span several files; it may fix a bug, implement a new feature, improve some documentation. The intent is that most of those extensions get integrated into Sage as soon as they are mature enough, with a typical life-cycle ranging from a few days to a couple months. In average 20 extensions are merged in each version of Sage (42 in Sage 5.0!), and more than 200 are under development.

Why do we want to share our experimental code

Here are our goals in using the Sage-Combinat queue for sharing patches:

• Preintegration
  • It is very common that an advanced feature needs some infrastructure support. For example, advanced Hopf algebras or representation theory features needs basic linear algebra stuff (eg: tensor product) which them self needs support from categories which them self may need support for optimized dynamic classes. Having a central repository for experimental code allows us for sharing several layer of dependant patch. In our (Nicolas and Florent) experience it is fairly common that during research, we end up improving dependant patches with more than four layers of dependencies.
• Pair programming (or more than pair!)
  • Many Sage-Combinat patches have several authors. We need an easy way to exchange patches (note that this is not specific to the Sage-Combinat project).
• Easy review even with many dependencies
  • As said in preintegration, we can have several layer of dependant patches. We need some tool to apply a bunch of patches (not necessarily in a stable/needs-review status), experiment with the code and launch the tests.
• Maturation
  • Due to the kind of computation we need (gluing algebra and combinatorics together), we have to be extra careful on the interface. Therefore, it is very common that we wait for a feature to be used several time before entering Sage. This is particularly true for infrastructure stuff.
• Overview of what's developed by who
  • Having a centralized place where all development is seen is a good tool for team coordination and code management. It also helps early detection of patch conflicts.
• Sharing code with beginner colleagues
  • The queue is also an easy way to distribute experimental code to non developer colleagues. The two commands
    • sage -combinat install
    • sage -combinat update
    need no mercurial or developer skills

== What are our constraints ==

• Vital necessity of supporting several versions of Sage at once
• For the convenience of the user, it is usually possible to use the sage-combinat patches with older versions of sage. The intent is only to temporarily support one or two older versions of sage (that is about one month old). Typical use case: a developer urgently needs the latest version of a patch for a software demonstration at a conference, but can't instantly upgrade because of a slow internet connection. There is no guarantee whatsoever; on occasion we do not support this when this causes technical difficulties.
• By nature, our calculations are transversal. Thus it would be hard to split Sage-Combinat in smaller chunks by subareas.

Some random questions

• linear order versus DAG (directed acyclic graph) of dependencies: what's easier to maintain ?

Foreseeable future

• More contributors
• Less overlap between patches as development goes from core to peripheral features