A Conversation with Badri Janakiraman about Hexagonal Rails
The recent conversation I’ve been having with Kent Beck, David Heinemeier Hansson about whether TDD is dead has spawned plenty of other conversations. One such exchange was with my colleague, Badri Janakiraman, one of our senior developers at Thoughtworks. Badri has over a decade of experience, working both in Ruby on Rails, but also in other development stacks with Java and .NET. For the last six years he’s been working in Studios, our product division, where much of his time has been with Mingle, our project collaboration tool that is a long running Rails project.
Since our discussions were so interesting, I thought it would be fun to capture them on video. So we scheduled some time together for a video chat, which I think you’ll find an interesting viewpoint into the role of Hexagonal Architecture in Ruby on Rails, and an useful complement to the “Is TDD Dead” discussion.
1: Active Record or Data Mapper?
5 June 2014
We introduce the notion of Hexagonal Architecture, and Badri describes the trade-offs between using Active Record or a Data Mapper to push the database outside the hexagon.
Badri began by explaining the notion of Hexagaonal Architecture, originally laid out by Alistair Cockburn. The main idea of this architecture is that your domain is a self-contained entity with no dependencies on external components. This allows you to reason about, and test, your application independently from those components. This leads to the question of what is external and what is internal? The descriptions of hexagonal architecture in Rails usually consider the database as external, but that isn't a given. The original article also considers driving your application through different forms of UI.
Martin drew the relationship between hexagonal architecture and the notion of a layered architecture. A common layering is some variation on having layers for UI, domain, and data source: with dependencies running UI → domain → data source. Hexagonal architecture changes the dependencies between domain and data source so your dependencies run: UI → domain ←︎ data source. When Martin wrote P of EAA, he described this approach as the Data Mapper pattern, which is in contrast to Active Record which ties the domain to the data source.
Martin asked Badri how he characterized the differences between Data Mapper and Active Record. He replied that it's all about the kind of application you need to build. In his early career he built applications with complex business logic such as leasing and insurance. When the domain has that kind of complexity you need to separate it from database concerns.
In the last six years or so, he's been working with rails on product applications with much simpler domain logic. For example Mingle is a project collaboration tool that helps you manage your backlog with a card wall metaphor. Since Thoughtworks's philosophy is that the tool should not constrain a team's process, we need lots of flexibility for the data, required user defined fields. The usual approaches for user-defined fields don't work very well, so we dynamically alter the relational schema to add custom fields. This has worked very well and ties the domain objects to the database, bringing the database into the heart of the application. Martin summarized that the theme here is that Data Mapper should be used when you want domain model and data source to evolve independently.
Badri continued by saying his pre-rails work involved cases where the application developers didn't have complete control over the database structure. This makes Data Mapper more attractive. Data Mapper gives you more isolation from the database, but it isn't complete isolation - you still need to take into account data management concerns, such as loading the same data with different domain model objects depending on the particular need.
Badri described how Mingle works with a few different databases (Postgres, Oracle, and at one time MySQL). The fact that the unit tests hit the database greatly helps portability across these databases - contrary to the usual notion that says you need isolation for this.
Martin summarized the trade-off between Data Mapper and Active Record by saying that the reasons to use Data Mapper are if you don't control the database (for instance when you are using an integration database) or if you have complex domain logic. But if these forces aren't in play, then Active Record is simpler (and hence better). We point out that the size of the database isn't a factor here: Mingle is a 50 KLOC app with over a hundred database tables, yet works well with Active Record.
Martin asked how long it takes for the tests to run. Badri replied that the tests run in around ten minutes or so with tests running in parallel with multiple browsers and databases. The team finds a ten minute commit suite is fast enough to maintain flow. They commit into a pending head.
2: Rails: Platform or Suite of Components?
12 June 2014
When working with a rich framework like Ruby on Rails you can treat it as with a platform or as a suite of components. Badri discusses the difference between these and we discuss what trade-offs go with the decison.
We started this conversation with Martin asking what the attitude towards Rails (and other similar frameworks) should be. Should we treat them as platforms, accepting them, warts and all? Or should we treat them a suite of components in which you pick the parts of the framework you like and leave out the parts that you don't. Badri responded by saying that while Rails has come quite a ways in being more modular, it is possible today to subset Rails more easily than to swap out whole components. For example, one could leave out Action Mailer in an application that doesn't need to either send or receive emails. One could pull in Action Support into a standard Ruby application and get a lot of the convenience methods and core-extensions that it provides. That said, one cannot, for example trivially replace the Rails request-response dispatch cycle with an evented system- like the one in Node.js for example. And this is because Rails is built to support a very specific kind of application. Martin chuckled, referring to the trope that Rails is a framework meant to build Basecamp.
Badri countered by saying that the sweet spot for Rails is a deeply webby application that talks to a relational datastore. In other words, your application has limited number of users and regardless of whether it is humans using your application through a web UI or other machines interacting with your application through a REST API, the web was not an accidental element. Badri further speculated that perhaps the problem is that people decide on using frameworks too early. Frameworks take away a lot of design options in return for a lot of convenience- and that means that perhaps we should understand how we would want to design our application and then pick a framework that doesn't prohibit those design choices rather than choosing a framework first and having it get in the way of how you would like to design the system
Martin replied that the choice to use the framework and design your application one way might have been made some time ago. That said, the framework designers might choose a different road and you might find yourself stuck in a corner. Martin then summarized our conversation so far by saying that if you are in the sweet spot for the framework, it makes sense to embrace it as a platform rather than spending the effort upfront to isolate oneself from possible future change. Badri agreed with this summary and said that some people find the additional layers they put in to be beneficial in guiding their design thoughts. That said, he personally tends to lie on the side of using Rails as a platform.
His personal preference for using a framework as a platform came from experiences in trying to circumvent frameworks in the past. Using the example of rich-client frameworks in .NET, He recalled the one time when he worked on an application where they attempted to not use two-way data binding. Due to how some members of the team felt the application should be tested, they tried to create a humble view by not using the data binding mechanisms. This led to a considerable mess because the framework was not designed to be used that way and we ended up writing far too much code and re-inventing major chunks of the framework. This is the kind of cost that needs to be understood by teams who wish to run against the grain of a framework. Perhaps the cost is sometimes worth it, but the fact that there is often a huge cost is unavoidable. He added that teams might wish to pick best of breed libraries and building a-la-carte applications if they find themselves wishing to make significant changes to a framework. Martin reiterated that the point is to make an informed choice to use a framework as a Platform or as Suite of Components driven by need- and not by blindly going full-embrace of either mode without having taken into consideration the costs and benefits of either approach.
Martin asked if there is value in frameworks making certain design choices for you and if that enables new members of the team to get started more quickly. Badri responded by saying that this is an advantage of frameworks in general - including ones such as Spring in the Java world. He claimed that he personally found the common usage patterns that people follow when using a framework to be more important- citing the example of how back in the Enterprise Java world, around 2001, almost every application would have it's own implementation of where transaction boundaries lay and how they were implemented. This was a something that he had to learn over and over in each application he worked on. In Rails, on the other hand, there is a common pattern of starting transactions and rolling them or committing them in a filter that wraps every request-response cycle. This sort of learning is useful to be able to carry from one Rails codebase to the next.
Martin called out for people not to push frameworks into places which are not their sweet spot. Badri added that this while this point is perfectly valid, Rails certainly does leave the door wide open for all sorts of design choices in places where the framework does not have an opinion. He illustrated this with a couple of examples from projects he was involved in. The first example is where they would fetch commit data from a version control system so that it could be shown in the context of other data in the application. Due to the needs of having the application work without this functionality and also the need to have data from multiple version control systems all be pulled into the application, they designed a port which defined the point of entry for this information into the system. This port became the single point for adapting all incoming data on by appropriate use of adapters for each kind of data that would arrive onto this port.
The second example was of two way communication where they built a Jabber chatroom bot that would pull in messages in team chat rooms that were relevant to stories that were being worked on, and attach them as a discussion thread to the story in the application that managed the backlog. Once again, because this functionality wasn't core to the application and they needed to let individual teams turn this functionality on or off as needed, they implemented a gateway that took care of all communication between the chat room and the core application which was completely oblivious as to whether this functionality was even configured. This gave them a good level of isolation for the core of the domain logic.
Martin summarized these two examples by saying that using Rails is not about avoiding avoiding isolation completely. He empasizes that you don't want to isolate yourself from your platform- but you do want to isolate yourself from external pieces. The distinction between what constitutes external and what constitues internal parts of your application brings us back a full circle to where we started our chat with Hexagonal Architecture pattern.