Our experience with using third-party libraries

NDepend is a tool that helps .NET developers write beautiful code. The project was started in April 2004. It is now used by more than 6 000 companies worldwide.

In more than a decade, many decisions were made, each with important consequences on the code base evolution process, sometime good consequences, sometime less good. Relentlessly dog-fooding (i.e using NDepend to analyze and improve the NDepend code base) helped us a lot to obtain more maintainable code, less bugs, and to improve the tool usability and features.

When it comes to working on and maintaining a large code base for several years, some of the most important decisions made are related to relying (or not) on some third-party libraries. Choosing whether or not to rely on a library is a double-edged sword decision that can, over time, bring a lot of value or cause a lot of friction. Ultimately users won’t make a difference between your bugs and third-party libraries bugs, their problems become your problems. Consider this:

• Sometimes the team just needs a fraction of a library and it may be more cost-effective, and more maintainable with time, to develop your own.
• Sometimes the licensing of a free or commercial library will prevent you from achieving your goals.
• Sometimes the library looks bright and shiny but becomes an obsolete project a few months later and precious resources will have to be spent maintaining others code, or migrating toward a trendy new library.
• Sometimes the library code and/or authors are so fascinating that you’ll be proud to contribute and be part of it.

Of course we all hope for the last case, and we had the chance to experience this a few times for real. Here are some libraries we’ve had great success with:

Mono.Cecil (OSS)

Mono.Cecil is an open source project developed by Jean-Baptiste Evain that can read and write data nested in .NET assemblies, including compiled IL code. NDepend has relied on Cecil for almost a decade to read IL code and other data. Both the library and the support provided are outstanding. The performance of the library is near optimal and all bugs reported were fixed in a matter of days or even hours. For our usage, the library is actually close to bug free. If only all libraries had this level of excellence…

DevExpress WinForm (Commercial)

NDepend has also relied on DevExpress WinForm for almost a decade to improve the UI look and feel. NDepend is a Visual Studio extension and DevExpress WinForm makes smooth visual integration with Visual Studio. Concretely, thanks to this library we achieved the exact same Visual Studio theming and skinning, docking controls a la Visual Studio, menus, bars and special controls like BreadCrumb to explore directories. We have never been disappointed with DevExpress WinForm. The bugs we reported were fixed quickly, it supports high DPI ratio and it is rock solid in production.

Microsoft Automatic Graph Layout MSAGL (OSS)

NDepend has relied on MSAGL for several years to draw all sorts of graphs of dependencies between code elements including Call Graphs, Class Inheritance Graphs, Coupling Graphs, Path and Cycle Graphs…  This library used to be commercial but nowadays OSS.  Here also the bugs we reported were fixed quickly, it supports high DPI ratio and it is perfectly stable in production.

NRefactory (OSS)

NDepend used to have a C# Code Query LINQ editor in 2012, a few years before Roslyn became RTM. We wanted to offer users a great editing experience with code completion and documentation tooltips on mouse-hovering. At that time NRefactory was the best choice and it proved with the years to be stable in production. Nowadays Roslyn would certainly be a better choice, but since our initial investment NRefactory still does the job well, we didn’t feel the need (yet) to move to Roslyn.

Here are a few things we prefer to keep in-house:

Licensing

While there are libraries for licensing, these are vital, sensitive topics that require a lot of flexibility with time, and we preferred to avoid delegating it. This came at the cost of plenty of development/support and significant levels of acquired expertise. Even taking into account that these efforts could have been spent on product features, we still don’t regret our choice.

The licensing layer is a cornerstone in our relation with our users community and it cannot suffer from any compromise. As a side remark, several times I observed that the cost of developing a solid licensing layer postponed promising projects to become commercial for a while.

Persistence

As most of application, NDepend persists and restores a lot of data, especially the large amount of data in analysis results. We could have used relational or object databases, but for a UI tool embedded in VS, the worst thing would be to slow down the UI and make the user wait. We decided only optimal performance is acceptable for our users, and optimal performance in persistence means using custom streams of bytes. The consequence of this decision is less flexibility: each time our data schema evolves, we need to put in extra effort to keep ascendant compatibility.

I underline that most of the time it is not a good idea to develop a custom persistence layer because of the amount of work and expertise required. But taken account our particular needs and goals, I think we took the right decision.

Production Logs

I explained here about our production logs system. We consider it an essential component to making NDepend a super-stable product. Here also, we could have used a third-party library. We capitalize on our own logging system because, year after year, we customized it with a plethora of production information, which was required to help fix our very own problems. We kept the system lightweight and flexible, and it still helps us improve the overall stability and correctness of our products.

Dependency Matrix and Treemap UI Components

These UI components were developed years ago and are still up to date. Both then and now, I believe there is no good third-party alternative that meets all our requirements in terms of layout and performance. A few times, we received propositions to buy those components, but we are not a component provider and don’t have plans for that.

In this post I unveiled a few core choices we made over the years. We hope this information will be useful for other teams.