Linking Modular Architecture to Development Teams

A Case Study in Mobile

Can a modular architecture improve software delivery? Yes! -but with some caveats. This article charts the journey of an enterprise who set out to shift their architecture to a more modular one in order to ease their growing pains. They found that modularity is a multifaceted solution that extends beyond architecture, into business lines of communication, team topologies and effective developer experience. By paying close attention to these factors, the enterprise was able to achieve significant uplifts in the delivery performance of their mobile applications.

13 June 2023


Photo of Matthew Foster

Matthew is a Technical Principal at Thoughtworks. Over his 14 years as a technologist, Matthew has spent his time building solutions and leading teams for businesses both large and small across retail, energy, telecommunications and Government.


This article will demonstrate the direct links between different mobile scaling issues, technical architecture and teams. At Thoughtworks we work with many large enterprises each presenting different problems and requirements when scaling their mobile presence. We identify two common problems seen in large enterprise mobile app development:

  1. A gradual lengthening of the time it takes to introduce new features to a market app
  2. Internal feature disparity arising from a lack of compatibility/reusability between in-house market apps

This article charts the journey one of our clients took when trying to address these issues. We tell the story of how their organisation had in the past, gravitated towards correct solutions, but was not able to see the expected benefits due to a misunderstanding of how those solutions were intrinsically linked.

We develop this observation by recounting how the same organisation was able to achieve a 60% reduction in average cycle time, an 18 fold improvement in development costs and an 80% reduction in team startup costs by shifting their Team Topologies to match a modular architecture while at the same time, investing in the developer experience.

Recognising the Signs

Despite the best of intentions, software often deteriorates over time, both in quality and performance. Features take longer to get to market, service outages become more severe and take longer to resolve, with the frequent result that those working on the product become frustrated and disenfranchised. Some of this can be attributed to code and its maintenance. However, placing the blame solely on code quality feels naive for what is a multifaceted issue. Deterioration tends to grow over time through a complex interplay of product decisions, Conway's law, technical debt and stationary architecture.

At this point, it seems logical to introduce the organisation this article is based around. Very much a large enterprise, this business had been experiencing a gradual lengthening of the time it took to introduce new features into their retail mobile application.

As a starter, the organisation had correctly attributed the friction they were experiencing to increased complexity as their app grew- their existing development team struggled to add features that remained coherent and consistent with the existing functionality. Their initial reaction to this had been to ‘just add more developers’; and this did work to a point for them. However, eventually it became apparent that adding more people comes at the expense of more strained communication as their technical leaders started to feel the increased coordination overhead. Hence the Two Pizza Team rule promoted at Amazon: any team should be small enough to be fed by two pizzas. The theory goes that by restricting how big a team can become, you avoid the situation where communication management takes more time than actual value creation. This is sound theory and has served Amazon well. However, when considering an existing team that has simply grown too big, there is a tendency towards 'cargo culting' Amazon’s example to try and ease that burden…

Limiting Cognitive Load

Indeed, the organisation was no exception to this rule: Their once small monolith had become increasingly successful but was also unable to replicate the required rate of success as it grew in features, responsibilities and team members. With looming feature delivery deadlines and the prospect of multiple brand markets on the horizon, they responded by splitting their existing teams into multiple smaller, connected sub-squads - each team isolated, managing an individual market (despite similar customer journeys).

This in fact, made things worse for them, as it shifted the communication tax from their tech leadership to the actual team itself, while easing none of their expanding contextual load. Realizing that communication and coordination was sapping an increasing amount of time from those tasked with actual value creation, our initial suggestion involved the idea of ‘cognitive load limitation’ outlined by Skelton & Pais (2019). This involves the separation of teams across singular complex or complicated domains. These seams inside software can be used to formulate the aforementioned ‘two pizza sized teams’ around. The result is much less overhead for each team: Motivation rises, the mission statement is clearer, while communication and context switching are shrunk down to a single shared focus. This was in theory a great solution to our client’s problem, but can actually be misleading when considered in isolation. The benefits from cognitive load limitation can only truly be realised if an application’s domain boundaries are truly well defined and consistently respected inside the code.

Domain Driven Discipline

Domain Driven Design (DDD) is useful for organising complex logic into manageable groups and defining a common language or model for each. However, breaking apart an application into domains is only part of an ongoing process. Keeping tight control of the bounded context is as important as defining the domains themselves. Examining our client’s application’s code we encountered the common trap of a clear initial investment defining and organising domain responsibilities correctly, only to have started to erode that discipline as the app grew. Anecdotal evidence from stakeholders suggested that perpetually busy teams taking shortcuts driven by urgent product requirements had become the norm for the team. This in turn had contributed to a progressive slowing of value delivery due to the accumulation of technical debt. This was highlighted further still by a measurable downtrend in the application’s Four Key Metrics as it became more difficult to release code and harder to debug issues.

Further warning signs of a poorly managed bounded context were discovered through common code analysis tools. We found a codebase that had grown to become tightly coupled and lacking in cohesion. Highly coupled code is difficult to change without affecting other parts of your system. Code with low cohesion has many responsibilities and concerns that do not fit within its remit, making it difficult to understand its purpose. Both these issues had been exacerbated over time as the complexity of each domain inside our client’s app had grown. Other indications came with reference again to cognitive load. Unclear boundaries or dependencies between domains in the application meant that when a change was made to one, it would likely involuntarily affect others. We noticed that because of this, development teams needed knowledge of multiple domains to resolve anything that might break, increasing cognitive load. For the organisation, implementing rigorous control of each domain-bounded context was a progressive step forward in ensuring knowledge and responsibility lay in the same place. This resulted in a limitation of the ‘blast radius’ of any changes, both in the amount of work and knowledge required. In addition, bringing in tighter controls in the accruing and addressing of technical debt ensured that any short term ‘domain-bleeds’ could be rejected or rectified before they could grow

Another metric that was missing from the organisation’s mobile applications was optionality of reuse. As mentioned earlier, there were multiple existing, mature brand market applications. Feature parity across those applications was low and a willingness to unify into a single mobile app was difficult due to a desire for individual market autonomy. Tight coupling across the system had reduced the ability to reuse domains elsewhere: Having to transplant most of an existing mobile app just to reuse one domain in another market brought with it high integration and ongoing management costs. Our utilisation of proper domain-bounded context control was a good first step to modularity by discouraging direct dependencies on other domains. But as we found out was not the only action we needed to take.

Domains that Transcend Apps

Scenario 1 - ‘The Tidy Monolith’

When viewed as a single application in isolation, simply splitting the app into domains, assigning a team, and managing their coupling (so as not to breach their bounded contexts) works very well. Take the example of a feature request to an individual application:

The feature request is passed to the app squads that own the relevant domain. Our strict bounded context means that the blast radius of our change is contained within itself, meaning our feature can be built, tested and even deployed without having to change another part of our application. We speed up our time to market and allow multiple features to be developed simultaneously in isolation. Great!

Indeed, this worked well in a singular market context. However as soon as we tried to address our second scaling problem- market feature disparity arising from a lack of reusability - we started to run into problems.

Scenario 2 - ‘The Next Market Opportunity’

The next step for the organization on its quest for modularity of domains was to achieve rapid development savings by transplanting parts of the ‘tidy monolith’ into an existing market application. This involved the creation of a common framework (aspects of which we touch on later) that allowed functionalities/domains to be reused in a mobile application outside its origin. To better illustrate our methodology, the example below shows two market applications, one in the UK, the other, a new app based out of the US. Our US based application team has decided that in addition to their US specific domains they would like to make use of both the Loyalty Points and Checkout domains as part of their application and have imported them.

For the organisation, this appeared to mean an order of magnitude development saving for their market teams vs their traditional behaviour of rewriting domain functionality. However, this was not the end of the story- In our haste to move towards modularity, we had failed to take into account the existing communication structures of the organisation that ultimately dictated the priority of work. Developing our previous example as a means to explain: After using the domains in their own market the US team had an idea for a new feature in one of their imported domains. They do not own or have the context of that domain so they contact the UK application team and submit a feature request. The UK team accepts the request and maintains that it sounds like “a great idea”, only they’re currently “dealing with requests from UK based stakeholders” so it's unclear when they will be able to get to the work...

We found that this conflict of interest in prioritising domain functionality limits the amount of reuse a consumer of shared functionality could expect - this was evident with market teams becoming frustrated at the lack of progress from imported domains. We theorized a number of solutions to the problem: The consuming team could perhaps fork their own version of the domain and orchestrate a team around it. However, as we knew already, learning/owning an entire domain to add a small amount of functionality is inefficient, and diverging also creates problems for any future sharing of upgrades or feature parity between markets. Another option we looked into was contributions via pull request. However this imposed its own cognitive load on the contributing team - forcing them to work in a second codebase, while still depending on support on cross team contributions from the primary domain team. For example, it was unclear whether the domain team would have enough time between their own market’s feature development to provide architectural guidance or PR reviews.

Scenario 3 - ‘Market Agnostic Domains’

Clearly the problem lay with how our teams were organised. Conway’s law is the observation that an organisation will design its business systems to mirror its own communication structure. Our previous examples describe a scenario whereby functionality is, from a technical standpoint modularised, however from an ownership standpoint is still monolithic: “Loyalty Points was created originally for the UK application so it belongs to that team”. One potential response to this is described in the Inverse Conway Maneuver. This involves altering the structure of development teams so that they enable the chosen technical architecture to emerge.

In the below example we advance from our previous scenario and make the structural changes to our teams to mirror the modular architecture we had previously. Domains are abstracted from a specific mobile app and instead are autonomous development teams themselves. When we did this, we noticed relationships changed between the app teams as they no longer had a dependency on functionality between markets. In their place we found new relationships forming that were better described in terms of consumer and provider. Our domain teams provided the functionality to their market customers who in turn consumed them and fed back new feature requests to better develop the domain product.

The main advantage this restructuring has over our previous iteration is the clarification of focus. Earlier we described a conflict of interest that occurred when a market made a request to change a domain originating from within another market. Abstracting a domain from its market changed the focus from building any functionality solely for the benefit of the market, to a more holistic mission of building functionality that meets the needs of its consumers. Success became measured both in consumer uptake and how it was received by the end user. Any new functionality was reviewed solely on the amount of value it brought to the domain and its consumers overall.

Focus on Developer Experience to Support Modularity

Recapping, the organisation now had a topological structure that supported modularity of components across markets. Autonomous teams were assigned domains to own and develop. Market apps were simplified to configuration containers. In concept, this all makes sense – we can plot how feedback flows from consumer to provider quite easily. We can also make high level utopian assumptions like: “All domains are independently developed/deployed” or “Consumers ‘just’ pull in whatever reusable domains they wish to form an application”.

In practice, however, we found that these are difficult technical problems to solve. For example, how do you maintain a level of UX/brand consistency across autonomous domain teams? How do you enable mobile app development when you are only responsible for part of an overall application? How do you allow discoverability of domains? Testability? Compatibility across markets? Solving these problems is entirely possible, but imposes its own cognitive load, a responsibility that in our current structure did not have any clear owner. So we made one!

A Domain to Solve Central Problems

Our new domain was categorised as ‘the platform’. The platform was essentially an all encompassing term we used to describe tooling and guidance that enabled our teams to deliver independently within the chosen architecture. Our new domain team maintains the provider/consumer relationship we have seen already, and is responsible for improving the developer experience for teams that build their apps and domains within the platform. We hypothesised that a stronger developer experience will help drive adoption of our new architecture.

But ‘Developer Experience’ (DX) is quite a non-specific term so we thought it important to define what was required for our new team to deliver a good one. We granularised the DX domain down to a set of necessary capabilities – the first being, Efficient Bootstrapping.

With any common framework there is an inevitable learning curve. A good developer experience aims to reduce the severity of that curve where possible. Sensible defaults and starter kits are a non-autocratic way of reducing the friction felt when onboarding. Some examples we defined for our platform domain:

We Promise that:

  • You will be able to quickly generate a new domain with all associated mobile dependencies, common UI/UX, Telemetry and CI/CD infrastructure in one command
  • You will be able to build, test and run your domain independently
  • Your domain will run the same way when bundled into an app as it does independently”

Note that these promises describe elements of a self-service experience within a developer productivity platform. We therefore saw an effective developer platform as one that allowed teams that were focused around end-user functionality to concentrate on their mission rather than fighting their way through a seemingly endless list of unproductive tasks.

The second necessary capability we identified for the platform domain was Technical Architecture as a Service. In the organisation, architectural functions also followed Conway’s law and as a result the responsibility for architecture decisions was concentrated in a separate silo, disconnected from the teams needing the guidance. Our autonomous teams, while able to make their own decisions, tended to need some aspect of ‘technical shepherding’ to align on principles, patterns and organisational governance. When we extrapolated these requirements into an on demand service we created something that looks like:

We Promise that:

  • The best practice we provide will be accompanied with examples that you can use or actual steps you can take
  • we'll maintain an overall picture of domain usage per app and when needed, orchestrate collaboration across verticals
  • The path to production will be visible and correct
  • We will work with you”

Note that these promises describe a servant leadership relationship to the teams, recognizing that everyone is responsible for the architecture. This is in contrast to what some might describe as command and control architectural governance policies.

One last point on the Platform Domain, and one worth revisiting from the previous example. In our experience, a successful platform team is one that is deeply ingrained with their customer’s needs. In Toyota lean manufacturing, “Genchi Genbutsu” roughly translates to “Go and see for yourself”. The idea being that by visiting the source of the problem and seeing it for yourself, only then can you know how to fix it. We learned that a team with the focus of improving developer experience must be able to empathise with developers that use their product to truly understand their needs. When we first created the platform team, we did not give this principle the focus it deserved, only to see our autonomous teams find their own way. This ultimately caused duplication of efforts, incompatibilities and a lack of belief in the architecture that took time to rectify.

The Results

We’ve told the story about how we modularised a mobile app, but how successful was it over time? Obtaining empirical evidence can be difficult. In our experience, having a legacy app and a newly architected app within the same organisation using the same domains with delivery metrics for both is a scenario that doesn’t come around too often. However luckily for us in this instance, the organisation was large enough to be transitioning one application at a time. For these results, we compare two functionally similar retail apps. One legacy with high coupling and low cohesion albeit with a highly productive and mature development team (“Legacy monolith”). The other, the result of the modular refactoring exercise we described previously - a well defined and managed bounded context but with ‘newer’ individual domain teams supporting (“Domain-bounded Context App”). Cycle time is a good measure here as it represents the time taken to ‘make’ a change in the code and excludes pushing an app to the store- A variable length process that App type has no bearing on.

Mobile App Type Cycle Time
Legacy Monolith17 days
Domain Bounded Context (Avg)10.3 days

Even when cycle time was averaged across all domain teams in our second app we saw a significant uplift versus the Legacy App with a less experienced team.

Our second comparison concerns optionality of re-use, or lack thereof. In this scenario we examine the same two mobile apps in the organisation. Again, we compare one requiring existing domain functionality (with no choice but to write it themselves) with our modular app (able to plug and play an existing domain). We ignore the common steps on the path to production since they have no impact on what we are measuring. Instead, we focus on the aspects within the control of the development team and measure our development process from pre-production ‘product sign off’ to dev-complete for a single development pair working with a designer full-time.

Integration Type Avg Development Time
Non-modular90 days
Modular5 days

The dramatically different figures above show the power of a modular architecture in a setting that has a business need for it.

As an aside, it is worth mentioning that these external factors we have excluded should also be measured. Optimising your development performance may reveal other bottlenecks in your overall process. For example, if it takes 6 months to create a release, and governance takes 1 month to approve, then governance is a comparatively small part of the process. But if the development timeline can be improved to 5 days, and it still takes 1 month to approve, then compliance may become the next bottleneck to optimise.

One other advantage not represented in the results above is the effect a team organised around a domain has on integration activities. We found autonomous domain teams naturally seconding themselves into market application teams in an attempt to expedite the activity. This, we believe, stems from the shift in focus of a domain squad whereby success of its domain product is derived from its adoption.

We discovered two concentric feedback loops which impact the rate of adoption. The outer, a good integration experience from the consumer of the domain (i.e. the app container). This is a developer-centric feedback loop, measured by how easily the consumer could configure and implement the domain as part of their overall brand-specific product offering. The inner, a good end user experience - how well the overall journey (including the integrated domain) is received by the consumer’s market customer. A poor consumer experience impacts adoption and ultimately risks insulating the domain team from the actual users of the capability. We found that domain teams which collaborate closely with consumer teams, and which have direct access to the end users have the fastest feedback loops and consequently were the most successful.

The final comparison worth mentioning is one derived from our Platform domain. Starting a new piece of domain functionality is a time consuming activity and adds to the overall development cost for functionality. As mentioned earlier, the platform team aims to reduce this time by identifying the pain points in the process and optimising them – improving the developer experience. When we applied this model to domain teams within our modular architecture we found an over 80% reduction in startup costs per team. A pair could achieve in an afternoon activities that had been estimated for the first week of team development!

Limitations

By now you should have quite a rosy picture of the benefits of a modular architecture on mobile. But before taking a sledgehammer to your ailing monolithic app, it's worth bearing in mind the limitations of these approaches. Firstly, and indeed most importantly, an architectural shift such as this takes a lot of ongoing time and effort. It should only be used to solve serious existing business problems around speed to market. Secondly, giving autonomy to domain teams can be both a blessing and a curse. Our platform squad can provide common implementations in the form of sensible defaults but ultimately the choices are with the teams themselves. Naturally, coalescing on platform requirements such as common UI/UX is in the interest of the domain squads if they wish to be incorporated/accepted into a market app. However, managing bloat from similar internal dependencies or eclectic design patterns is tricky. Ignoring this problem and allowing the overall app to grow uncontrolled is a recipe for poor performance in the hands of the customer. Again, we found that investment in technical leadership, in conjunction with robust guardrails and guidelines helps to mitigate this problem by providing architecture/design oversight, guidance and above all communication.

Summary

To recap, at the start of this article we identified two significant delivery problems exhibited in an organisation with a multi app strategy. A lengthening of the time it took to introduce new features into production and an increasing feature disparity between other similar in house applications. We demonstrated that the solution to these problems lies not in a single strategy around technical architecture, team structure or technical debt, but in a simultaneously evolving composite of all those aspects. We started by demonstrating how evolving team structures to support the desired modular and domain-centric architecture improves cognitive and contextual load, while affording teams the autonomy to develop independently of others. We showed how a natural progression to this was the elevation of teams and domains to be agnostic of their originating application/market, and how this mitigated the effects of Conway’s law inherent with an application monolith. We observed that this change allowed a consumer/provider relationship to naturally occur. The final synchronous shift we undertook was the identification and investment in the ‘platform’ domain to solve central problems that we observed as a consequence of decoupling teams and domains.

Putting all these aspects together, we were able to demonstrate a 60% reduction in cycle time averaged across all modular domains in a market application. We also saw an 18 fold improvement in development cost when integrating modular domains to a market app rather than writing from scratch. Furthermore, the focus on engineering effectiveness allowed our modular architecture to flourish due to the 80% reduction in startup costs for new domains and the ongoing support the ‘platform team’ provided. In real-terms for our client, these savings meant being able to capitalise on market opportunities that were previously considered far too low in ROI to justify the effort - opportunities that for years had been the uncontested domains of their competitors.

The key takeaway is that a modular architecture intrinsically linked to teams can be highly beneficial to an organisation under the right circumstances. While the results from our time with the highlighted organisation were excellent, they were specific to this individual case. Take time to understand your own landscape, look for the signs and antipatterns before taking action. In addition, do not underestimate the upfront and ongoing effort it takes to bring an ecosystem like that which we have described together. An ill considered effort will more than likely cause more problems than it solves. But, by accepting that your situation will be unique in scope and thus resisting the pull of the ‘cargo cult’: Focusing on empathy, autonomy and lines of communication that enable the architecture at the same time, then there is every reason you could replicate the successes we have seen.


Acknowledgments

Special thanks Carl Nygard, Tim Cochran and Greg Davis for all your suggestions and critical feedback. Thanks also to Rob Horn for persuading me to put my thoughts in order and actually write this article. Finally, thanks to Martin Fowler for your guidance.

Significant Revisions

13 June 2023: Published