Summarized using AI

Modularizing Rails Monoliths One Bite at a Time

Marc Reynolds • October 26, 2023 • Boulder, CO • Talk

In the talk titled "Modularizing Rails Monoliths One Bite at a Time" presented at Rocky Mountain Ruby 2023, Marc Reynolds discusses the challenges and solutions associated with managing complexity in Rails monoliths as they scale. He highlights that while many developers turn to microservices for relief, they often face increased complexity instead. An alternative approach, known as the Modular Monolith, is proposed, emphasizing modularization to enforce code boundaries without the disadvantages of microservices. Key points from the talk include:

  • The Problem of Spaghetti Code: As applications evolve, codebases often degrade into complex 'spaghetti code' where maintainability suffers, leading to bloated controllers and models, which complicates development and increases bug occurrences.
  • Packwerk Gem: Marc introduces the Packwerk gem, a tool that aids in maintaining clear boundaries within the codebase through static code analysis to detect privacy violations. By structuring code into cohesive packages, teams can enhance clarity and manage dependencies effectively.
  • Phased Approach to Modularization:
    • Understanding Feature Sets: Collaborating with product leaders to map out features aids in defining package structures.
    • Creating Initial Packages: The initial phase involves simply moving code into designated packages based on feature sets without restructuring, ensuring speed and urgency.
    • Declaring Dependencies: The next phase focuses on identifying and resolving dependency violations while maintaining cohesion within packages.
    • Establishing Public Interfaces: Finally, teams must create stable public interfaces that outline clear contracts for interactions between packages, allowing internal iterations without impacting external dependencies.
  • Long-term Maintenance: Ensuring continuous integration practices, monitoring for privacy violations, and keeping documentation up to date are crucial for maintaining a robust modular structure.
  • Conclusion: Marc concludes that with the right practices and tools, applications can regain clarity and maintain a rewarding development experience, steering clear of the pitfalls of microservices. By following a phased approach, teams can modularize their monoliths effectively, fostering collaboration and easing onboarding for new developers.

Modularizing Rails Monoliths One Bite at a Time
Marc Reynolds • October 26, 2023 • Boulder, CO • Talk

Rocky Mountain Ruby 2023 - Modularizing Rails Monoliths One Bite at a Time by Marc Reynolds

As Rails monoliths grow, coupling becomes increasingly difficult to manage. Many have reached for hope in microservices but instead found higher complexity. The Modular Monolith approach is a proven, lightweight alternative that offers the benefits of enforced boundaries without being cumbersome. This talk proposes a phased approach to refactoring toward this style using the packwerk gem.

Rocky Mountain Ruby 2023

00:00:13.759 Ready to talk about spaghetti code today. Does anybody here write spaghetti code?
00:00:19.560 All right, good. Intentionally, all right?
00:00:26.199 Cool. All right, everybody else can go because we're going to talk about spaghetti code today.
00:00:32.000 So as Spike mentioned, I'm an engineering manager at Doximity where I build software to help doctors do their jobs better every day.
00:00:39.800 I am from New Mexico, where we fly around in wicker baskets that shoot fire.
00:00:46.440 It's a pretty cool place, and I have a cute little tortoise. His name is Bunny Rabbit—big thanks to my daughter for naming animals confusing names.
00:00:53.160 So if anyone's looking for a low-maintenance pet, in about 60 years he's going to outlive me, so he might be available for you if you're interested in a tortoise.
00:01:06.000 Also, I've been working on software for the medical industry for just about 20 years, covering all sorts of areas.
00:01:11.799 From management software to scheduling software for hospitals, operating room monitoring software, and now tools for doctors' everyday processes.
00:01:17.240 It’s a cool industry to be in.
00:01:24.040 So we're going to start with a story from Doximity.
00:01:29.360 At the end of 2016, a few of us at Doximity started up a new Rails application, a microservice that was isolated from our main Rails monolith.
00:01:36.560 This isolation gave us freedom to iterate and do basically whatever we wanted.
00:01:42.840 Our goal was to establish clear boundaries of code ownership, and our spec suite ran in under two minutes—it was amazing.
00:01:49.640 This Greenfield application became increasingly important and grew into one of the most critical applications in our ecosystem.
00:01:56.719 The number of contributors to our Git repo went from about a dozen to nearly 170.
00:02:02.479 This growth became a bit confusing.
00:02:08.560 Our previously tidy Rails app became really difficult to work with, as Active Record classes and relationships were being used in unintended ways.
00:02:14.680 We constantly battled N+1 queries, and queries would run for 15 minutes, often with no idea why.
00:02:21.400 Developers had a hard time figuring out where the code belonged.
00:02:29.280 New developers would come in and struggle to figure out where things were.
00:02:35.120 So the question was, should we just fork off another microservice and repeat the process?
00:02:43.599 Could we ever get back to that bliss of a Greenfield app again?
00:02:49.840 This is a common experience with Rails applications, leading me to formulate a pattern of degeneration.
00:02:56.360 First, we start to see our controllers getting bloated, prompting us to refactor to thin controllers and thick models.
00:03:02.400 If we've felt that pain before, we tend to move on to models that become bloated, identifying God models and refactoring to thin models and thick services.
00:03:09.640 Initially, this sounds great until we end up with a giant bucket of intertwined general services that everyone uses, but no one understands.
00:03:17.000 Does anybody else have a codebase that sounds like that? Exactly.
00:03:22.040 We started seeing bugs every time we changed our microservices just a little.
00:03:26.199 We entered lockdown mode, where we were really afraid to change anything because we weren’t sure if our specs covered all the creative usages the other teams were utilizing.
00:03:35.120 The good news is that our apps, which we love so much, don't have to stay in a state of confusion.
00:03:42.840 By taking some measured, intentional steps, we can dig ourselves out of the hole of complexity and coupling.
00:03:50.039 We can leverage simple tools to enforce code boundaries, leaning into modularization.
00:03:56.400 By discussing method contracts, which we don’t often do, we can return to multiple lands of Greenfield bliss.
00:04:04.960 Packwerk is a tool that has helped us take an iterative approach to regain that Greenfield bliss.
00:04:10.920 I’ll provide some links to Packwerk and other resources later on.
00:04:16.000 Essentially, Packwerk is a static code analysis tool that detects private code access by modules outside of a package.
00:04:23.360 It prescribes a folder structure that’s somewhat similar to the standard Rails application but includes privacy significance for certain folders.
00:04:31.200 We moved all of our code from the standard MVC structure and distributed it into several packages that were highly cohesive but loosely coupled.
00:04:39.120 For example, in the public directory, we’ve created packs for sales, reviews, and social.
00:04:46.639 Everything within the public directory is accessible; if it’s not in that directory, other packages can’t access it.
00:04:53.760 Let’s take a quick example. We have two packages: a sales pack and a review pack.
00:05:01.040 The sale model is not in the public directory, so it's not public.
00:05:06.000 However, the review model in the reviews package has a method called sales rolling 30 that leverages the sale model.
00:05:15.040 This is a violation, as it’s reaching into a file located outside of the public directory.
00:05:22.640 What are some possible problems with this? This seems like a simple issue that happens frequently.
00:05:29.680 What if the sales team's application maintainers never indexed the created-at field?
00:05:36.600 It might work fine in development, possibly in QA, and maybe even in pre-production—but then you go to production.
00:05:43.960 Suddenly, the method takes 15 seconds to complete, creating issues for your production users.
00:05:50.000 What if the sales team wants to remove the finalized method or change its meaning?
00:05:56.800 That’s going to require a conversation with the review team.
00:06:03.240 What if the sales team never reviewed this PR?
00:06:09.600 Does this mean that every PR needs to be reviewed by every team?
00:06:15.400 At one point, I was reviewing 85 PRs a month from other teams accessing our methods.
00:06:22.080 We still had issues slip through because not everyone can review every single PR.
00:06:30.000 We can utilize Packwerk to check for these problems; you don’t have to review every single PR.
00:06:38.560 We have several command line tools within the Packwerk gem.
00:06:45.040 The first tool is Packwerk check, which scans the entire codebase to detect privacy violations, like the one we just covered.
00:06:53.760 If we run Packwerk check on our application, we would see a message indicating that we have a privacy violation.
00:07:01.440 It will inform us that there's a violation with the sale model in the sales package.
00:07:08.480 This is a check that can be run in your CI process.
00:07:16.600 Having perfect little packages of code might seem amazing and be a great goal, but when you're staring at a 200,000 line codebase, it's tough.
00:07:24.640 It can be hard to know where to start, gain buy-in from stakeholders, and get motivated.
00:07:31.680 This phased approach can be critical in maintaining momentum in a huge project.
00:07:38.720 Here's a sneak peek of the whole talk: we’ll examine what we actually have in our codebase.
00:07:45.680 Next, we’ll chunk files into packages, declare package dependencies, start writing public interfaces, and then maintain it.
00:07:52.240 Let’s start with the basics, which isn’t even in our codebase.
00:08:00.560 First, I’d recommend getting a lay of the land to figure out your true feature set.
00:08:07.680 Work with your product leaders to understand how they view groups of features in your codebase.
00:08:14.080 Understanding how the business thinks about roadmaps and planning and the language they use is crucial.
00:08:24.000 Design groups of features and give them names for the basis of your package structure.
00:08:30.000 This approach will likely differ from what the engineers would consider.
00:08:38.640 We used a process called event storming.
00:08:45.000 We brought together our product leaders, tech leads, and directors over a couple of days to discuss the events a user experiences.
00:08:52.560 We put sticky notes on a window, took pictures, and these became the folders, akin to the directory structures I mentioned earlier.
00:08:59.199 The next phase involves using those feature sets from the planning phase to create folders and move code into them.
00:09:04.960 When our team entered this phase, we moved classes wholesale without changing any namespaces or making any refactorings.
00:09:10.640 This helped us move quickly without getting bogged down in decisions.
00:09:16.680 It might be worthwhile to examine whether the code is being used.
00:09:23.200 If it's not being executed, just delete it and reduce the code you have to manage.
00:09:30.480 Packwerk uses YAML files to declare dependencies in these package.yaml files.
00:09:38.560 Inside, you'll set up an array of dependencies.
00:09:44.240 We held off on setting up our dependencies at this stage; we’ll get to that in a little bit.
00:09:51.360 However, we added some metadata to our files declaring the owning team, Slack channel, and the code reviewer group.
00:09:58.960 This helped in two ways: it provided a way for anyone exploring the code to know precisely where to ask questions.
00:10:05.280 With 20 or 30 packages, this information is critical.
00:10:11.040 We used this code owner piece to autogenerate our GitHub code owners file, which is very convenient.
00:10:18.160 We moved nearly everything into packages.
00:10:25.920 My perspective here was to think about what would need to be removed if I were to delete a feature.
00:10:32.520 Everything related to that feature—the models, views, controllers, rake tasks, and database migrations—moved into that package.
00:10:39.679 This leads to a highly cohesive package where everything inside is related, while maintaining less coupling between packages.
00:10:47.280 The only code we left outside packages was application configuration code that didn’t correlate to any single feature set, such as our DB schema.
00:10:53.920 We also added a README to each package directory, which isn't required by Packwerk but is a good practice.
00:10:59.440 These README files outlined the purpose of each package and explained why certain dependencies exist.
00:11:04.960 This clarity becomes essential when assumptions change or when we need to consider removing a dependency.
00:11:11.760 As we go through this phase, everything will show as a privacy violation when you run Packwerk check.
00:11:18.640 You might see a staggering number of violations because nothing is in the public directory.
00:11:25.800 This might seem daunting to look at, but Packwerk provides a solution by allowing us to generate to-do files.
00:11:31.760 This lists all the files and all the violations, indicating to Packwerk we are aware and will address them later.
00:11:41.760 We can run Packwerk update to-do, which performs the same code scanning Packwerk check does, but creates a YAML file.
00:11:46.960 It's like list-based development—the clearer we can get about our objectives, the more manageable our task.
00:11:55.440 When we ran the Packwerk update to-dos command for every PR that moved files into packages, we generated massive to-do files.
00:12:01.440 You’ll likely have numerous violations; that’s okay, as it indicates we want to increase the number of files in our package directory.
00:12:08.640 Phase two focuses on declaring dependencies between packages.
00:12:16.080 We’ll start removing dependency violations while leaving privacy violations for now.
00:12:22.560 In this phase, we draw coarse-grained lines between our packages.
00:12:31.680 For example, in our package YAML file, we declare that our sales package depends on reviews and inventories.
00:12:38.240 Fixing these dependency violations will allow us to focus on the privacy violations which remain.
00:12:45.920 As we declare package dependencies in package YAML files, we can run Packwerk validate.
00:12:53.920 This ensures we don't have any circular dependencies, which can be embedded in your CI process.
00:13:02.640 If there’s a circular dependency between the sales package and the inventory package, it will be flagged.
00:13:11.200 Eliminating those circular dependencies is essential.
00:13:17.600 You may discover tough circular dependencies that require rethinking how your code relates.
00:13:24.160 This is when you realize there is a significant amount of coupling between features.
00:13:31.280 In our case, it helped to categorize packages.
00:13:38.160 We created infrastructure packages for tasks like JSON parsing, messaging, and instrumentation—working with services outside of our application.
00:13:46.800 Next, we had common tooling packages.
00:13:55.760 If we had a base Active Record class or functionality related to the frontend that didn’t relate to specific features, we’d throw it into the common tools package.
00:14:02.960 The feature packages contained specific business logic with minimal dependencies.
00:14:10.240 Both common tooling and infrastructure did not have outgoing dependencies and thus remained independent of the features.
00:14:19.360 This approach somewhat aligns with Domain-Driven Design (DDD) architecture; think of it as DDD light.
00:14:26.960 However, you may find dependencies that cannot be simply rectified by breaking code into smaller packages.
00:14:34.960 Imagine a user profiles package and a social package where users make comments and likes.
00:14:42.560 We may encounter a circular dependency if user interactions require execution whenever users are deleted.
00:14:50.160 In such cases, consider implementing a Publish-Subscribe (pub/sub) architecture.
00:14:57.600 Social can subscribe to user profiles, signaling events without needing to know about the underlying package.
00:15:04.240 We used the Rails Event Store gem, a lightweight event bus—an ideal choice for lightweight eventing.
00:15:11.920 Phase three involves making tough decisions.
00:15:19.000 At this point, we have many privacy violations but declared all dependencies and resolved any circular ones.
00:15:26.560 Now, we need to create public interfaces for other packages.
00:15:33.680 These public interfaces are stable, serving as a contract for other teams, requiring discussions for changes.
00:15:42.560 However, teams can freely iterate on internal code without extensive code reviews.
00:15:51.520 This internal structure once again becomes that amazing development platform.
00:15:59.680 Here's an example of creating a public interface.
00:16:06.480 We emphasized extensive documentation; every public interface needed clear comments about its functionality.
00:16:12.960 This clarifies what the public interface does, doesn’t do, its requirements, and potential side effects.
00:16:20.000 We also annotated parameters to clearly articulate expected inputs.
00:16:27.440 Duck typing is great, but it can lead to unexpected issues, so we were specific about types.
00:16:34.000 We're cautious about returning Active Record objects as it can lead to leakage of private methods.
00:16:41.840 We implemented a to_data method that converts Active Record objects into plain old Ruby objects.
00:16:48.920 This abstraction allows us to return data objects without exposing the underlying Active Record structure.
00:16:56.640 This reduces potential N+1 issues since it streamlines data access.
00:17:02.720 After this, any returned data object minimizes the risk of mutating data or querying child relations.
00:17:09.680 This level of abstraction brings peace of mind when other teams utilize our public interfaces.
00:17:17.120 Regarding specs, expect many violations since we often do unconventional things.
00:17:25.120 Factory Bot is invaluable but often leads to confusion between production and testing environments.
00:17:31.360 If your models are kept internal to a package, specs outside can't assume how data is structured.
00:17:38.560 Think of specs as integration tests.
00:17:47.040 Refactored specs should avoid Factory Bot and instead use concrete methods for generating data.
00:17:55.040 Using actual creation methods creates a stable foundation for expectations.
00:18:02.200 At this point, we should ensure a high level of cohesion within our packages, while maintaining low coupling.
00:18:09.680 Packages should change independently, while everyone is aware of what belongs where.
00:18:16.800 New developers will appreciate this clarity; they can quickly find what they're searching for.
00:18:23.760 Phase four is crucial as it determines whether all the hard work pays off.
00:18:31.200 Here’s some advice on maintaining your packages.
00:18:39.120 First, ensure to run these checks in CI.
00:18:46.560 It is impractical to review every single pull request, especially after all this work.
00:18:53.360 Keep an eye on privacy violations or circular dependency issues.
00:19:01.680 Eventually, those YAML files should be deleted; they should not reappear unless you agree to take on that technical debt.
00:19:08.480 Watch out for everything ending up public.
00:19:15.920 Sometimes developers forget the reasons behind public exposure, leading to a convoluted mess.
00:19:23.520 Keep conversations active regarding the necessity of public abstractions.
00:19:30.200 Ensure other teams can access what they need from your packages.
00:19:37.680 Keep utilizing the domain model, and reconsider the way functionalities are organized.
00:19:44.400 One way to speed things up is by exposing Active Record classes, leading back to the threats of N+1 issues.
00:19:52.080 Returning plain old Ruby objects is reliable.
00:19:59.760 Documentation must stay fresh.
00:20:07.040 Utilize README files to explain why dependencies exist.
00:20:13.840 Reassess your assumptions regularly.
00:20:21.440 New developers will benefit from scanning through README files for clear insights.
00:20:30.080 Continue gathering feedback continuously, not just once.
00:20:36.480 Engagement with your support team helps clarify how external teams interact.
00:20:42.560 For example, our SRE noted the challenge of connecting stack traces to specific packages.
00:20:49.440 We discussed using namespace patterns to clarify these interactions.
00:20:57.760 Lastly, enjoy the process.
00:21:05.280 We’ve all experienced the excitement of using Rails and the freedom that comes with it.
00:21:13.280 If you manage to create clear, enjoyable package structures, it’s an excellent position to be in.
00:21:21.120 Don’t hesitate to explore new implementations that maintain these contracts.
00:21:30.000 Ensure each package maintains clarity and ease of use.
00:21:38.560 Here's that phased approach summarized again.
00:21:47.040 This is a long haul for extensive codebases with many contributors.
00:21:54.960 Understand what you have before jumping into assumptions regarding package setups.
00:22:02.000 Gather input from product teams to align language.
00:22:09.760 Create a package layout that reflects the business's perspective.
00:22:17.920 Establish package dependencies and consider why features need to be closely coupled.
00:22:26.960 Cruise your way toward clarity in your codebase.
00:22:36.320 Reassure yourself that the dream of clarity in your code is achievable.
00:22:45.040 Your application doesn’t have to remain confusing; you needn't fall back on microservices.
00:22:52.400 Remember, Rails engines are another option but may not offer the same privacy scanning.
00:23:01.120 Find a balance between leaving a giant mess and starting fresh.
00:23:09.200 Let’s keep pushing for that Monday morning excitement, building excellent software that tackles hard problems.
00:23:16.640 Thank you!
Explore all talks recorded at Rocky Mountain Ruby 2023
+10