Microservices are often described as small, loosely coupled applications that follow the UNIX philosophy of “doing one thing well.” They have also been related to the Single Responsibility Principle, the first of the five principles making up SOLID. A microservices-based architecture is typically constructed around a set of common patterns. This set of patterns is actually consistent with all of the SOLID principles when thought of at the architectural rather than the class/module level. In this article, we’ll gain an understanding of each of the SOLID principles and how they relate to microservices.
A SOLID Architecture?
I have spent a significant portion of the last three years speaking and writing about the SOLID principles of object-oriented design. I spent the first year teaching SOLID much as expressed by Robert C. “Uncle Bob” Martin in his foundational book, Agile Software Development: Principles, Patterns, and Practices. About the same time I was embarking on a reintroduction to and exploration of functional programming, spending a significant amount of time with Clojure. Retraining my mind for functional thinking while simultaneously teaching the SOLID principles resulted in a perfect thought storm in my mind, resulting in the following question:
Is there an overlap between functional programming and the SOLID principles?
The thought journey that followed led me to start giving the talk entitled “Functional SOLID” on August 25, 2012. That same month I began publishing a four-part series of articles by the same name. In both of these presentations of my ideas, I related a couple of foundational memes:
Functional programming, especially within the Clojure programming language, provides wonderful constructs for building programs consistent with the SOLID principles.
The SOLID principles actually transcend all of programming, regardless of the particular programming paradigm (structured, object-oriented, functional) employed.
It is this second meme that provides the impetus for this particular article. At the conclusion of both presentations, I refer to Rich Hickey’s seminal Strange Loop keynote, “Simple Made Easy”. In his presentation, Hickey decries our continual conflation of the ideas of simplicity and ease by tracing the origins of both words:
Simple, from the Latin simplex, meaning “one fold or braid” (opposed to complex, meaning “many folds or braids,”) is an objective concept. In software we can relate it to the degree of interleaving of concerns in software components.
Easy, from the Latin adjacens, means “to lie near.” While “hard” does not trace to a root meaning “to lie far,” we can still understand that “easy” is a relative concept. What lies near, or is easy to me, is not necessarily easy or near to you. In software we can relate it to the programming languages, paradigms, framweworks, technologies, etc. with which we are intimately familiar.
This analysis led me to restate the SOLID principles in terms of another Rich Hickey word, complectedness, or the degree to which software concerns are interleaved:
- Single Responsibility Principle
Complecting responsibilities leads to rigid and/or fragile design.
- Open-Closed Principle
Complecting concretions of an abstraction in such a way that new concretions adversely affect existing, working concretions is problemmatic.
- Liskov Substitution Principle
Reuse via inheritance is dangerous. It often complects entities not in a true “is-a” relationship, which leads to non-substitutability.
- Interface Segregation Principle
Don’t complect clients with uninteresting operations by complecting unrelated groups of operations in a single entity!
- Dependency Inversion Principle
Transitive dependency leads to transitive complectedness!
As you can see, we’ve now stated the principles independently of any programming-paradigm specific language. It is this restatement that cemented in my mind the idea that SOLID actually transcends all of software engineering — software engineering principles that are universally true, regardless of the context in which they are applied. So, we can easily walk these principles up the abstraction ladder into the world of architecture.
With businesses built around software now disrupting multiple industries that appeared to have stable leaders, the need has emerged for enterprises to create “software factories” built around the following principles:
Streaming customer feedback directly into rapid, iterative cycles of application development;
Horizontally scaling applications to meet user demand;
Compatibility with an enormous diversity of clients, with mobility (smartphones, tablets, etc.) taking the lead;
Continuous delivery of value, shrinking the cycle time from concept to cash.
Infrastructure has taken the lead in adapting to meet these needs with the move to the cloud, and Platform as a Service (PaaS) has raised the level of abstraction to a focus on an ecosystem of applications and services. However, most applications are still developed as if we’re living in the previous generation of both business and infrastructure: the monolithic application. Microservices — small, loosely coupled applications that follow the Unix philosophy of "doing one thing well" — represent the application development side of enabling rapid, iterative development, horizontal scale, polyglot clients, and continuous delivery. They also enable us to scale application development and eliminate long term commitments to a single technology stack.
I won’t belabor the introduction to microservices anymore, as a wealth of reading is already available all over the web. As a jumping off point, I invite the reader to dive into James Lewis' and Martin Fowler’s excellent coverage of the topic.
Let’s now get to the heart of the matter: how do the patterns associated with a microservices architecture overlap with the SOLID principles? Let’s walk through each, briefly relating them in their natural context, and then swinging them into our microservices discussion.
Single Responsibility Principle
The Single Responsibility Principle (SRP) is stated by Martin as “each software module should have one and only one reason to change” [SRP]. Of all of the SOLID principles, the SRP is the one I’ve most often seen cited in the context of microservices.
One common thread between how Martin relates SOLID and microservices is change. Change in software is inevitable and constant. Requirements are realized as responsibilities doled out to various software modules. Requirements change leads to changes in responsibilities. If we couple responsibilities in a single module, then change to one responsibility can affect another unrelated responsibility simply due to its location. In other words, change one thing, sometimes another unrelated thing breaks. Risk goes up; change velocity goes down.
A monolithic architecture, no matter how modular on the inside, couples responsibilities together. Change cycles are coupled, increasing the risk associated with frequent deployments. Effective continuous delivery is far more difficult, as the release management process reimposes the waterfall process on the agile development team. If we instead separate architectural responsibilities into different microservices, we can decouple those change cycles, thus decreasing the risks associated with frequent deployments. Continuous delivery becomes more easily attainable.
The most common technique I’ve seen applied to decomposing a monolith into microservices is the bounded context from Domain-Driven Design. We identify discrete business capabilities, each of which owns and governs its own discrete segment of the overall data model for an organization. A microservice implements each business capability, encapsulating its data segment behind an often RESTful API. Overlaps between the capabilities (e.g. a shipping service and ordering service will both likely have the notion of customer, likely governed by a customer service) are realized by mappings in higher-order microservices or by utilizing hypermedia.
The Open-Closed Principle (OCP), first coined by Bertrand Meyer [OOSC], states that “software entities should be open for extension, but closed for modification.” Again we relate this principle to change. We should be able to change what a module does as software requirements change, but we should be able to do so without modifying any existing, working code.
At face value this looks impossible. How can we change the behavior of a module without changing its code? The key is in how we define the facade of the module, thinking at the appropriate level of abstraction.
Let’s draw an example from Java’s standard library.
What if my client code is provided an instance of
java.util.HashMap, and I instead want sorted keys?
I would need to not only provide an instance of
java.util.TreeMap to my client, but I would also need to change all of the existing references.
If I instead refer to the map abstraction as
java.util.Map (a Java interface), then I can provide my client with the new
Map type without changing any code.
By utilizing the appropriate module facade, we can decouple an abstraction from the its larger set of derivative behaviors.
What is our microservice facade? The API of course! As long as a given microservice continues to fulfill the contract expressed by its API, it should be possible to swap in new behaviors without changing any existing client code. This becomes supremely important when we consider the term of our commitment to a particular technology stack. Monolithic architectures are not closed to this particular type of modification, and the risk of incorporating new technology into an existing monolith can be very high. Microservices drastically reduce the risk associated with experimenting, even in production, with new technology stacks, and increase our ability to use the right tool for the job.
Another important technique enabled by the open-closed nature of microservices is polyglot persistence. By encapsulating the data store technology used for a particular business capability behind its facade (e.g. a recommendations service is very amenable to graph databases), we can hide the presence of that data store behind a microservice API. This enables us to both experiment with and utilize various data stores in advantageous contexts without polluting the overall service ecosystem with the semantics of each store.
Liskov Substitution Principle
The Liskov Substitution Principle (LSP) was born the same year as Meyer’s OCP, written down by Barbara Liskov.
The LSP is concerned with types and subtypes, focused on the idea that “subtypes must be substitutable for their base types.”
In object-oriented terms, drawing again from the Java language, if a class
extends from a parent class or
implements a parent interface, we should be able to use that class in the context of any code expecting an instance of the parent.
If at any time that code context exhibits aberrant behavior, we have violated the LSP with our class.
Extending the idea of object-oriented inheritance to logical architecture is a bit of a stretch, but let’s give it a try. We’ll start by again considering the microservice’s facade, or its API. From the client’s perspective, the API represents the “base type” for our microservice. So long as any microservice we swap in properly fulfills this API, we can say it’s consistent with the LSP.
It’s unlikely that we’ll often substitute different implementations of the same API at runtime, and it’s unclear to me what a child microservice might look like. However, consider the case of services that implement the same API, but that must implement different business rules or policies given the legal jurisdiction governing the data. Further, consider that regulatory compliance dicates that those services actually are deployed and run in the same geographic location governed by that legal jurisdiction. We could implement each instance of this API as a separate microservice and deploy each of them in the appropriate geography. From the client’s perspective, the substitution would be transparent (thus abiding by the LSP), and the “polymorphic” substitution could be performed by another higher-order microservice or global site-selection mechanism.
Interface Segregation Principle
The Interface Segregation Principle (ISP) is stated in Martin’s book as “clients should not be forced to depend on methods they do not use.” Martin introduces the concept of so-called “fat interfaces,” or interfaces whose method set is not cohesive. One can divide their method sets into multiple groups, each group serving a different set of interested clients. The primary reason for seeking to separate these groups into different modules is to prevent change driven by one set of clients from affecting other distinct groups of clients.
API’s implemented via monolithic architectures cannot abide by the ISP. Adding or improving capabilities to serve one group of clients must involve minimally a redeployment of all of the capabilities affecting all clients. More likely, a lengthy regression test phase will also be required, as we must ensure that these additional or improved capabilities have not damaged the system’s other capabilities.
Microservices, when designed well around bounded contexts, also abide by the ISP, as we enforce a hard boundary between interfaces by separating them into discrete, independently deployable units.
Dependency Inversion Principle
The Dependency Inversion Principle (DIP) tells us that “abstractions should not depend upon details. Details should depend upon abstractions.” Stated another way, “high-level modules should not depend on low-level modules.” Our abstractions, or higher-level modules, are what codify the important business knowledge inherent in a body of software, whereas our details, or lower-level modules, represent the mechanical recipes for carrying them out. One of the promises of the other principles is the ability to “swap out” the details beneath the abstractions when it becomes advantageous. However, when our higher-level modules have direct dependency on our lower-level modules, swapping out details often causes the abstraction itself to have to change. “Absurd” is Uncle Bob’s description of this situation.
The DIP typically deals with this scenario by defining service interfaces for each module. If a module requires services that are not relevant to its bounded context, rather than implementing them itself or directly delegating to a dependency, it instead declares a signature for that service within its service interface. This interface then becomes a secondary abstraction expressing all of the collaboration a module intends to do. Possible collaborators then cooperate with the module by implementing its service interface. In this way, they become dependent on the module, rather than the module becoming dependent on the collaborator!
In a microservices architecture, the DIP finds its realization in the API Gateway pattern. An API Gateway acts as a single point of entry into a microservices architecture for a given client. It plays a multi-faceted role in serving the diverse clients (i.e. disparate mobile device platforms) of the architecture by:
reducing the chattiness of the network by reducing the number of services consulted;
performing protocol translation (e.g. AMQP to HTTP) when a particular protocol is not well supported by the client;
aggregating service responses concurrently to reduce response latency;
transforming service responses to service the needs of specific devices, screen sizes, and use cases.
The important thing to note is that the “API” defined by the API Gateway is owned by the client in much the same way as the service interface is owned by a higher-level module. In this way, we invert the dependency between clients and the microservices themselves. Consult “Optimizing the Netflix API” for a fantastic example of this architectural pattern.
Experience has taught me that ease is often cheap but illusory, but that simplicity is a pearl of great price. Microservices are not easy, but they are simple. One of the reasons for their simplicity is what I see as their strong compatibility with the SOLID principles, not only of object-oriented design, but of all of software engineering. By resisting the temptation to interleave distinct business capabilities, we retain the ability to develop and deploy them in an agile manner. I hope you’ve found some value in this article, and I even hope you’ve found some things with which you disagree. Please sound off in the blogosphere, the Twitterverse, or wherever suits your fancy.
[SRP] Martin, Robert C. “The Single Responsibility Principle.” http://blog.8thlight.com/uncle-bob/2014/05/08/SingleReponsibilityPrinciple.html