Software Development

Software Development
(In a small-scale research context)

It is not hard to argue that the creation of software is the greatest human cultural achievement to date. But writing software well is not easy and much sophisticated methodology has been proposed for software development, primarily addressing the needs of large software companies and enterprise-scale systems. Certainly: once software development becomes the task of teams, and systems become larger than what one person can remember confidently, failure is virtually guaranteed if the task can't be organized in a structured way.

But our work often does not fit this paradigm, because in the bioinformatics lab the requirements change quickly. The reason is obvious: most of what we produce in science are one-off solutions. Once one analysis runs, we publish the results, and we move on. There is limited value in doing an analysis over and over again. However, this does not mean we can't profit from applying the basic principles of good development principles. Fortunately that is easy. There actually is only one principle.

Make implicit knowledge explicit.

Everything else follows.

 

Collaborate

Making project goals explicit and making progress explicit are crucial, so that everyone knows what's going on and what their responsibilities are. Collaboration these days is distributed, and online. Here is a list of options you ought to have tried. :

• Schedule regular face-to-face meetings. If you can't be in the same room, Google hangout's may work (up to ten people). Old-time developers often use IRC chat rooms.
• A wiki such as our Student Wiki is obviously a good way to structure, share and collaboratively edit information.
• Software sharing and code collaboration is best managed through a github repository. As a version control system, it is an essential part of all software development.
• Dynamic coauthoring capabilities are offered by Etherpad installations such as the one hosted by WikiMedia.
• You have probably collaborated with Google Drive text, spreadsheet or presentation documents before. These have superseded many other online offerings.
• A robust system for shared file storage is Dropbox.
• Trello appears to be a nice tool to distribute work-packages and keep up to date with discussions, especially if your "team" is distributed.
• I like Kanbanery as a structured To-Do list for my own time-management, but it can also be adapted to project workflows.

There is a lot of turnover with all of these tools and it's important to backup your important data to alternate locations. Offerings may go out of date, or get absorbed by other services. Are there collaboration tools you like that I haven't mentioned? Let me know ...

Plan

The planning stage involves defining the goals and endpoints of the project. We usually start out with a vague idea of something we would like to achieve. We need to define:

• where we are;
• where we want to be;
• and how we will get there.

For an example of a plan, refer to the 2015 BCB420 Class Project. There, we lay out a plan in three phases: Preparation, Implementation and Results. This is generic, the preparation phase implies an analysis of the problem, which focusses on what will be accomplished, independent of how this will be done. The results of the analysis can be a requirements document (here is a link to the ABP Requirements template - a template to jumpstart such a document) or a less formal collection of goals.

The most important achievement of the plan is to break down the project into manageable parts and define the Milestones that characterize the completion of each part.

 

Design

SPN (Structured Process Notation) is one way to define dataflow diagrams for data-driven analysis in bioinformatics.

In this example SPN icons are integrated to describe a workflow that annotates protein structure domains at the residue level.

In the design phase, we focus on the architecture of the system that fulfils the requirements, as well as the algorithms.

By architecture we mean the components, their interfaces and behaviour. Typically this will involve some modelling and there are different ways to model a system.

• Structural modelling describes the components and interfaces. The components are typically pieces of software, the interfaces are "contracts" that describe how information passes from one piece to another. Structural models include the Data model that captures how data reflects reality and how reality changes the data in our system;
• Behaviour modelling describes the state changes of our system, how it responds to input and how data flows through the system. In data-driven analysis, a data flow diagram may capture most of what is important about the system.
  • SPN is a data flow diagram method that I find particularly suited to the kind of code that we often write in bioinformatics: integration of data, transformation, and analytics, all passing through well defined intermediate stages. Read more about it here.

Typically, several different types of models may contribute to understanding a system; in practice dataflow diagrams may be particularly well suited for the workflow centric systems that we commonly encounter in bioinformatics.

Whichever modelling method is adopted, make sure that it is explicit. Many aspects of the widely used UML standard suffer from very poor information-design decisions; this can lead to hard-to-understand graphics, that obfuscate rather than clarify the issues. That, of course, is cargo cult design.

The design objectives have to be clearly articulated. There are many ways to achieve the same end result, but good design will create systems that are correct, robust, extensible and maintainable. A very good starting point for the design is a proper separation of concerns.

 

Develop

In the development phase, we actually build our system. It is a misunderstanding if you believe most time will be spent in this phase. Designing a system well is hard. Building it, if it is well designed, is easy. Building it if it is poorly designed is probably impossible.

A number of development methodologies and philosophies have been proposed, and they go in and out of fashion. In this course we will work with a conjunction of TDD (Test Driven Development) and Literate programming.

Literate Programming

Literate programming is an idea that software is best described in a natural language, focussing on the logic of the program, i.e. the why of code, not the what. The goal is to ensure that model, code, and documentation become a single unit, and that all this information is stored in one and only one location. The product should be consistent between its described goals and its implementation, seamless in capturing the process from start (data input) to end (visualization, interpretation), and reversible (between analysis, design and implementation).

In literate programming, narrative and computer code are kept in the same file. This source document is typically written in Markdown or LaTeX syntax and includes the programming code as well as text annotations, tables, formulas etc. The supporting software can weave human-readable documentation from this, or tangle executable code. Literate programming with both Markdown and LaTex is supported by R Studio and this makes the R Studio interface a useful development environment for this paradigm. While it is easy to edit source files with a different editor and process files in base R after loading the Sweave() and Stangle() functions or the knitr package. In our context here we will use R Studio because it conveniently integrates the functionality we need.

For exercises on knitr, RMarkdown and LaTex, follow this link.

Test Driven Development

TDD is meant to ensure that code actually does what it is meant to do. In practice, we define our software goals and devise a test (or battery of tests) for each. Initially, all tests fail. As we develop, the tests succeed. As we continue development

• we think carefully about how to break the project into components and structure them (units). These more or less do one, thing, one thing only, and don't have side-effects;
• we discipline ourselves to watch out for unexpected input, edge- and corner cases and unwarranted assumptions;
• we can be confident that later changes do not break what we have done earlier - because our tests keep track of the behaviour.

For an exercise in Test Driven Development, follow this link.

Typically testing is done at several levels:

• During the initial development phases unit testing continuously checks the function of the software units of the system.
• As the code base progresses, code units are integrated and begin interacting via their interfaces. These interfaces can be specified as "contracts" that define the conditions and obligations of an interaction. Typically, a contract will define the precondition, postcondition and invariants of an interaction. These can be verified by tests.
• Final tests verify the code, and validate its correct execution- just like a positive control in a lab experiment.

Finally, testing supports maintenance. When you find a bug, write a test that would have discovered it, then fix the bug. Also, be mindful that you may have made the same type of error elsewhere in your code. Search for these cases, write tests and fix them too.

 

Version Control

Must.

• Software Carpentry Git Introduction
• Git tutorials
• Introduction to GitHub

 

Fail Safe or Fail Fast?

Testing for correct input is a crucial task for every function, and R especially goes out of its way to coerce input to the type that is needed. This is making functions fail safe. Do consider the opposite philosophy however: "fail fast", to produce fragile code. You must test whether input is correct, but a good argument can be made that incorrect input should not be fixed, but the function should stop the program and complain loudly and explicitly about what went wrong. This - once again - makes implicit knowledge explicit, it helps the caller of the function to understand how to pass correct input, and it prevents code from executing on wrong assumptions. In fact, failing fast may be the real fail safe.

 

Code

Here is a small list of miscellaneous best-practice items for the phase when actual code is being written:

• Be organized. Keep your files in well-named folders and give your file names some thought.
• Use version control: git or svn.
• Use an IDE (Integrated Development Environment). Syntax highlighting and code autocompletion are nice, but good support for debugging, especially stepping through code and examining variables, setting breakpoints and conditional breakpoints are essential for development. For Rdevelopment, the R Studio environment provides syntax highlighting and a symbolic debugger.
• Design your code to be easily extensible and only loosely coupled. Your requirements will change frequently, make sure your code is modular and nimble to change with your requirements.
• Design reusable code. This may include standardized interface conventions and separating options and operands well.
• DRY (Don't repeat yourself): create functions or subroutines for tasks that need to be repeated. Or, seen from the other way around: whenever you find yourself repeating code, it's time to delegate that to a function instead. Incidentally, that also holds for database design: whenever you find yourself intersecting tables with similar patterns, perhaps it's a good idea to generalize one of the tables and merge the data from the other into it.
• KISS (Keep it simple): resist the temptation for particularly "elegant" language idioms and terse code.
• Comment your code. I can't repeat that often enough. Code is read very much more often than it is written. Unfortunately (for you) the one most likely to have to read and understand your convoluted code is you yourself, half a year later. So do yourself the favour to explain what you are thinking. Don't comment on what the code does - that is readable from the code itself - but why you do something the way you do.
• Be consistent.

Coding style

It should always be your goal to code as clearly and explicitly as possible. R has many complex idioms, and it being a functional language that can generally insert functions anywhere into expressions, it is possible to write very terse, expressive code. Don't do it. Pace yourself, and make sure your reader can follow your flow of thought. More often than not the poor soul who will be confused by a particularly witty use of the language will be you, yourself, half a year later. There is an astute observation by Brian Kernighan that applies completely:

"Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it."

A number of guides for R coding style are available ... expand here to see what I recommend (resp. require).

• Use informative and specific filenames for code sources; give them the extension .R
• Give your sources headers stating purpose, author, date and version information, and note bugs and issues.
• Give your functions headers that describe purpose, arguments (including required datatypes), and return values. Callers should be able to work with the function without having to read the code.
• Use lots of comments. Never describe what the code does, but explain why.
• Use separators (# --- SECTION -----------------) to structure your code.
• Indent comment hashes to align with the expressions in a block.
• Use only <- for assignment, not =
• ...but do use = when passing values into the arguments of functions.
• Don't use <<- (global assignment) except in very unusual cases. Actually never.
• Use the concise camelCaseStyle for variable names, never use the confusing.dot.style or the rambling separating_with_underscores_style.
• Define parameters at the beginning of the code, use all caps variable names (MAXWIDTH). Never have "magic numbers" appear in your code.
• In mathematical expressions, always use parentheses to define priority. Never rely on convention. (( 1 + 2 ) / 3 ) * 4
• Always separate operators and arguments with spaces.
• Never separate function names and the brackets that enclose argument lists.
• Don't abbreviate argument names. You can, but you shouldn't.

• Try to limit yourself to ~80 characters per line.
• Always use braces {}, even if you write single-line if statements and loops.
• Always use a space after a comma, and never before a comma.
• Always explicitly return values from functions, never rely on returning the last expression.
• Use spaces to align repeating parts of code, so errors become easier to spot.
• Don't repeat code. Use functions instead.
• Explicitly assign values to crucial arguments, even if you think you know that the value is the default.
• Never reassign reserved words.
• Don't use c as a variable name since c() is a function.
• Don't call your data frames df since df() is a function.
• Don't use semicolons, and don't write more than one command on a line.
• Don't use attach().
• It's safer to use for (i in seq(along=x)) {...} not for (i in 1:length(x)) {...} because of an unwanted loop if x == NULL
• If possible, do not grow data structures dynamically, but create the whole structure with NULL values, then assign values to its elements. This is much faster.

Specific naming conventions I like

isValid, hasNeighbour ... Boolean variables

findRange(), getLimits() ... simple function names (verbs!)

initializeTable() ... not initTab()

node ... for one element; nodes ... for more elements

nPoints ... for number-of

isError ... not isNotError: avoid double negation

---

Consider using the formatR package for consistent code.

And seriously - XKCD: Code Quality

Code Reviews

Regularly meet to review other team member's code. Keep the meetings short, and restrict to no more than 300 lines at a time.

1. The code author walks everyone through the code and explains the architecture and the design decisions. It should not be necessary to explain what the code does - if it is necessary, that points to opportunities for improvement.
2. Briefly consider improvements to coding style as suggestions but don't spend too much time on them (don't create a "Bicycle Shed" anti-pattern) - style is not the most important thing about the review. Be constructive and nice - you should encourage your colleagues, not demotivate them.
3. Spend most of the time discussing architecture: how does this code fit into the general lay of the land? How would it need to change if its context changes? Is it sufficiently modular to survive? What does it depend on? What depends on it? Does it apply a particular design pattern? Should it? Or has it devolved into an anti-pattern?
4. Focus on tests. What is the most dangerous error for the system integrity that the code under review could produce. Are there tests that validate how the code deals with this? Are there tests for the edge cases and corner cases? This is the best part about the review: bring everyone in the room on board of the real objectives of the project, by considering how one component contributes to it.
5. Finally, what's your gut feeling about the code: is there Code Smell? Are there suboptimal design decisions that perhaps don't seem very critical at the moment but that could later turn into inappropriate technical debt? Perhaps some refactoring is indicated; solving the same problem again often leads to vastly improved strategies.

Overall be mindful that code review is a sensitive social issue, and that the primary objective is not to point out errors, but to improve the entire team.

 

Deploy and Maintain

In our context, deployment may mean a single run of discovery and maintenance may be superfluous as the research agenda moves on.

But this does not mean we should ignore best practice in scientific software development: simple, but essential aspects like using version control for our code, using IDEs, writing test cases for all code functions etc. These aspects are very well covered in the open source Software Carpentry project and courses. Free, online, accessible and to the point. Go there and learn:

• Software Carpentry

 

 

Notes

 

Further reading and resources

Practice

• git cheat sheet
• 19 git tips for everyday use

Concepts

Greg Wilson et al. on the excellent principles of the Software Carpentry workshops:

Wilson et al. (2014) Best practices for scientific computing. PLoS Biol 12:e1001745. (pmid: 24415924)

[ PubMed ] [ DOI ]

• Software design
• Software pattern
• Software development process
• Software architecture
• Portal:Software testing

UML: The Positive Spin

Architecture modeling. A quite useful overview of systems modeling, part of the Microsoft Visual Studio documentation.

• Kim Waldén and Jean-Marc Nerson: Seamless Object-Oriented Software Architecture: Analysis and Design of Reliable Systems, Prentice Hall, 1995.

Sandve et al. (2013) Ten simple rules for reproducible computational research. PLoS Comput Biol 9:e1003285. (pmid: 24204232)

[ PubMed ] [ DOI ]

Article in Nature Biotechnology; note that successful here is meant to imply widely used. David Baker's Rosetta package is not mentioned, for example. Nevertheless: good insights in this.

Altschul et al. (2013) The anatomy of successful computational biology software. Nat Biotechnol 31:894-7. (pmid: 24104757)

[ PubMed ] [ DOI ]

Peng (2011) Reproducible research in computational science. Science 334:1226-7. (pmid: 22144613)

[ PubMed ] [ DOI ] Computational science has led to exciting new developments, but the nature of the work has exposed limitations in our ability to evaluate published findings. Reproducibility has the potential to serve as a minimum standard for judging scientific claims when full independent replication of a study is not possible.