- For object-oriented design we follow the SOLID principles. For microservice design we propose developers follow the “IDEALS”: interface segregation, deployability (is on you), event-driven, availability over consistency, loose-coupling, and single responsibility.
- Interface segregation tells us that different types of clients (e.g., mobile apps, web apps, CLI programs) should be able to interact with services through the contract that best suits their needs.
- Deployability (is on you) acknowledges that in the microservice era, which is also the DevOps era, there are critical design decisions and technology choices developers need to make regarding packaging, deploying and running microservices.
- Event-driven suggests that whenever possible we should model our services to be activated by an asynchronous message or event instead of a synchronous call. Availability over consistency reminds us that more often end users value the availability of the system over strong data consistency, and they’re okay with eventual consistency.
- Loose-coupling remains an important design concern in the case of microservices, with respect to afferent (incoming) and efferent (outgoing) coupling. Single responsibility is the idea that enables modeling microservices that are not too large or too slim because they contain the right amount of cohesive functionality.
In 2000 Robert C. Martin compiled the five principles of object-oriented design listed below. Michael Feathers later combined these principles in the SOLID acronym. Since then, the SOLID principles for OO design have been described in books and became well-known in the industry.
- Single responsibility principle
- Open/closed principle
- Liskov substitution principle
- Interface segregation principle
- Dependency inversion principle
A couple of years ago, I was teaching microservice design to fellow developers when one of the students asked, “Do the SOLID principles apply to microservices?” After some thought, my answer was, “In part.”
Months later, I found myself searching for the fundamental design principles for microservices (and a catchy acronym to go with it). But why would such a question be important?
We, as an industry, have been designing and implementing microservice-based solutions for over six years now. During this time, an ever-increasing number of tools, frameworks, platforms, and supporting products have established an incredibly rich technology landscape around microservices.
This landscape can make a novice microservice developer dizzy with the many design decisions and technology choices they can face in just one microservice project.
In this space, a core set of principles can help developers to aim their design decisions in the right direction for microservice-based solutions.
Although some of the SOLID principles apply to microservices, object orientation is a design paradigm that deals with elements (classes, interfaces, hierarchies, etc.) that are fundamentally different from elements in distributed systems in general, and microservices in particular.
Thus, we propose the following set of core principles for microservice design:
- Interface segregation
- Deployability (is on you)
- Availability over consistency
- Loose coupling
- Single responsibility
The principles don’t cover the whole spectrum of design decisions for microservices-based solutions, but they touch the key concerns and success factors for creating modern service-based systems. Read on for an explanation of these principles applied to microservices — the much-needed microservice “IDEALS.”
The original Interface Segregation Principle admonishes OO classes with “fat” interfaces. In other words, instead of a class interface with all possible methods clients might need, there should be separate interfaces catering to the specific needs of each type of client.
The microservice architecture style is a specialization of the service-oriented architecture, wherein the design of interfaces (i.e., service contracts) has always been of utmost importance. Starting in the early 2000s, SOA literature would prescribe canonical models or canonical schemas, with which all service clients should comply. However, the way we approach service contract design has changed since the old days of SOA. In the era of microservices, there is often a multitude of client programs (frontends) to the same service logic. That is the main motivation to apply interface segregation to microservices.
Realizing interface segregation for microservices
The goal of interface segregation for microservices is that each type of frontend sees the service contract that best suits its needs. For example: a mobile native app wants to call endpoints that respond with a short JSON representation of the data; the same system has a web application that uses the full JSON representation; there’s also an old desktop application that calls the same service and requires a full representation but in XML. Different clients may also use different protocols. For example, external clients want to use HTTP to call a gRPC service.
Instead of trying to impose the same service contract (using canonical models) on all types of service clients, we “segregate the interface” so that each type of client sees the service interface that it needs. How do we do that? A prominent alternative is to use an API gateway. It can do message format transformation, message structure transformation, protocol bridging, message routing, and much more. A popular alternative is the Backend for Frontends (BFF) pattern. In this case, we have an API gateway for each type of client — we commonly say we have a different BFF for each client, as illustrated in this figure.
Deployability (is on you)
For virtually the entire history of software, the design effort has focused on design decisions related to how implementation units (modules) are organized and how runtime elements (components) interact. Architecture tactics, design patterns, and other design strategies have provided guidelines for organizing software elements in layers, avoiding excessive dependencies, assigning specific roles or concerns to certain types of components, and other design decisions in the “software” space. For microservice developers, there are critical design decisions that go beyond the software elements.
As developers, we have long been aware of the importance of properly packaging and deploying software to an appropriate runtime topology. However, we have never paid so much attention to the deployment and runtime monitoring as today with microservices. The realm of technology and design decisions that here we’re calling “deployability” has become critical to the success of microservices. The main reason is the simple fact that microservices dramatically increase the number of deployment units.
So, the letter D in IDEALS indicates to the microservice developer that they are also responsible for making sure the software and its new versions are readily available to its happy users. Altogether, deployability involves:
- Configuring the runtime infrastructure, which includes containers, pods, clusters, persistence, security, and networking.
- Scaling microservices in and out, or migrating them from one runtime environment to another.
- Expediting the commit+build+test+deploy process.
- Minimizing downtime for replacing the current version.
- Synchronizing version changes of related software.
- Monitoring the health of the microservices to quickly identify and remedy faults.
Achieving good deployability
Automation is the key to effective deployability. Automation involves wisely employing tools and technologies, and this is the space where we have continuously seen the most change since the advent of microservices. Therefore, microservice developers should be on the lookout for new directions in terms of tools and platforms, but always questioning the benefits and challenges of each new choice. (Important sources of information have been the ThoughtWorks Technology Radar and the Software Architecture and Design InfoQ Trends Report.)
Here is a list of strategies and technologies that developers should consider in any microservice-based solution to improve deployability:
- Containerization and container orchestration: a containerized microservice is much easier to replicate and deploy across platforms and cloud providers, and an orchestration platform provides shared resources and mechanisms for routing, scaling, replication, load-balancing, and more. Docker and Kubernetes are today’s de facto standards for containerization and container orchestration.
- Service mesh: this kind of tool can be used for traffic monitoring, policy enforcement, authentication, RBAC, routing, circuit breaker, message transformation, among other things to help with the communication in a container orchestration platform. Popular service meshes include Istio, Linkerd, and Consul Connect.
- API gateway: by intercepting calls to microservices, an API gateway product provides a rich set of features, including message transformation and protocol bridging, traffic monitoring, security controls, routing, cache, request throttling and API quota, and circuit breaking. Prominent players in this space include Ambassador, Kong, Apiman, WSO2 API Manager, Apigee, and Amazon API Gateway.
- Serverless architecture: you can avoid much of the complexity and operational cost of container orchestration by deploying your services to a serverless platform, which follows the FaaS paradigm. AWS Lambda, Azure Functions, and Google Cloud Functions are examples of serverless platforms.
- Monitoring tools: with microservices spread across your on-premises and cloud infrastructure, being able to predict, detect, and notify issues related to the health of the system is critical. There are several monitoring tools available, such as New Relic, CloudWatch, Datadog, Prometheus, and Grafana.
- Log consolidation tools: microservices can easily increase the number of deployment units by an order of magnitude. We need tools to consolidate the log output from these components, with the ability to search, analyze, and generate alerts. Popular tools in this space are Fluentd, Graylog, Splunk, and ELK (Elasticsearch, Logstash, Kibana).
- Tracing tools: these tools can be used to instrument your microservices, and then produce, collect, and visualize runtime tracing data that shows the calls across services. They help you to spot performance issues (and sometimes even help you to understand the architecture). Examples of tracing tools are Zipkin, Jaeger, and AWS X-Ray.
- DevOps: microservices work better when devs and ops teams communicate and collaborate more closely, from infrastructure configuration to incident handling.
- Blue-green deployment and canary releasing: these deployment strategies allow zero or near-zero downtime when releasing a new version of a microservice, with a quick switchback in case of problems.
- Infrastructure as Code (IaC): this practice enables minimal human interaction in the build-deploy cycle, which becomes faster, less error-prone, and more auditable.
- Continuous delivery: this is a required practice to shorten the commit-to-deploy interval yet keeping the quality of the solutions. Traditional CI/CD tools include Jenkins, GitLab CI/CD, Bamboo, GoCD, CircleCI, and Spinnaker. More recently, GitOps tools such as Weaveworks and Flux have been added to the landscape, combining CD and IaC.
- Externalized configuration: this mechanism allows configuration properties to be stored outside the microservice deployment unit and easily managed.
The microservice architecture style is for creating (backend) services that are typically activated using one of these three general types of connectors:
- HTTP call (to a REST service)
- RPC-like call using a platform-specific component technology, such as gRPC or GraphQL
- An asynchronous message that goes through a queue in a message broker
The first two are typically synchronous, HTTP calls being the most common alternative. Often, services need to call others forming a service composition, and many times the interaction in a service composition is synchronous. If instead, we create (or adapt) the participating services to connect and receive messages from a queue/topic, we’ll be creating an event-driven architecture. (One can debate the difference between message-driven and event-driven, but we’ll use the terms interchangeably to represent asynchronous communication over the network using a queue/topic provided by a message broker product, such as Apache Kafka, RabbitMQ, and Amazon SNS.)
An important benefit of an event-driven architecture is improved scalability and throughput. This benefit stems from the fact that message senders are not blocked waiting for a response, and the same message/event can be consumed in parallel by multiple receivers in a publish-subscribe fashion.
The letter E in IDEALS is to remind us to strive for modeling event-driven microservices because they are more likely to meet the scalability and performance requirements of today’s software solutions. This kind of design also promotes loose-coupling since message senders and receivers — the microservices — are independent and don’t know about each other. Reliability is also improved because the design can cope with temporary outages of microservices, which later can catch up with processing the messages that got queued up.
But event-driven microservices, also known as reactive microservices, can present challenges. Processing is activated asynchronously and happens in parallel, possibly requiring synchronization points and correlation identifiers. The design needs to account for faults and lost messages — correction events, and mechanisms for undoing data changes such as the Saga pattern are often necessary. And for user-facing transactions carried over by an event-driven architecture, the user experience should be carefully conceived to keep the end-user informed of progress and mishaps.
Availability over Consistency
The CAP theorem essentially gives you two options: availability XOR consistency. We see an enormous effort in the industry to provide mechanisms that enable you to choose availability, ergo embrace eventual consistency. The reason is simple: today’s end users will not put up with a lack of availability. Imagine a web store during Black Friday. If we enforced strong consistency between stock quantity shown when browsing products and the actual stock updated upon purchases, there would be significant overhead for data changes. If any service that updates stock were temporarily unreachable, the catalog could not show stock information and checkout would be out of service! If instead, we choose availability (accepting the risk of occasional inconsistencies), users can make purchases based on stock data that might be slightly out-of-date. One in a few hundred or thousand transactions may end up with an unlucky user later getting an email apologizing for a cancelled purchase due to incorrect stock information at checkout time. Yet, from the user (and the business) perspective, this scenario is better than having the system unavailable or super slow to all users because the system is trying to enforce strong consistency.
Some business operations do require strong consistency. However, as Pat Helland points out, when faced with the question of whether you want it right or you want it right now, humans usually want an answer right now rather than right.
Availability with eventual consistency
For microservices, the main strategy that enables the availability choice is data replication. Different design patterns can be employed, sometimes combined:
- Service Data Replication pattern: this basic pattern is used when a microservice needs to access data that belongs to other applications (and API calls are not suitable to get the data). We create a replica of that data and make it readily available to the microservice. The solution also requires a data synchronization mechanism (e.g., ETL tool/program, publish-subscribe messaging, materialized views), which will periodically or trigger-based make the replica and master data consistent.
- Command Query Responsibility Segregation (CQRS) pattern: here we separate the design and implementation of operations that change data (commands) from the ones that only read data (queries). CQRS typically builds on Service Data Replication for the queries for improved performance and autonomy.
- Event Sourcing pattern: instead of storing the current state of an object in the database, we store the sequence of append-only, immutable events that affected that object. Current state is obtained by replaying events, and we do so to provide a “query view” of the data. Thus, Event Sourcing typically builds upon a CQRS design.
A CQRS design we often use at my workplace is shown in the figure next. HTTP requests that can change data are processed by a REST service that operates on a centralized Oracle database (this service uses the Database per Microservice pattern nonetheless). The read-only HTTP requests go to a different backend service, which reads the data from an Elasticsearch text-based data store. A Spring Batch Kubernetes cron job is executed periodically to update the Elasticsearch store based on data changes executed on the Oracle DB. This setup uses eventual consistency between the two data stores. The query service is available even if the Oracle DB or the cron job is inoperative.
In software engineering, coupling refers to the degree of interdependence between two software elements. For service-based systems, afferent coupling is related to how service users interact with the service. We know this interaction should be through the service contract. Also, the contract should not be tightly coupled to implementation details or a specific technology. A service is a distributed component that can be called by different programs. Sometimes, the service custodian doesn’t even know where all the service users are (often the case for public API services). Therefore, contract changes should be avoided in general. If the service contract is tightly coupled to the service logic or technology, then it is more prone to change when the logic or technology needs to evolve.
Services often need to interact with other services or other types of components thus generating efferent coupling. This interaction establishes runtime dependencies that directly impact the service autonomy. If a service is less autonomous, its behavior is less predictable: in the best-case scenario, the service will be as fast, reliable, and available as the slowest, least reliable, and least available component it needs to call.
Loose coupling strategies for services
The letter L in IDEALS prompts us to be attentive to coupling for services and therefore microservices. Several strategies can be used and combined to promote (afferent and efferent) loose coupling. Examples of such strategies include:
- Point-to-point and publish-subscribe: these building block messaging patterns and their variations promote loose coupling because senders and receivers are not aware of each other; the contract of a reactive microservice (e.g., a Kafka consumer) becomes the name of the message queue and the structure of the message.
- API gateway and BFF: these solutions prescribe an intermediary component that deals with any discrepancies between the contract of the service and the message format and protocol that the client wants to see, hence helping to uncouple them.
- Contract-first design: by designing the contract independently of any existing code we avoid creating APIs that are tightly coupled to technology and implementation.
- Hypermedia: for REST services, hypermedia helps frontends to be more independent of service endpoints.
- Façade and Adapter/Wrapper patterns: variations of these GoF patterns in microservice architectures can prescribe internal components or even services that can prevent undesirable coupling to spread across a microservice implementation.
- Database per Microservice pattern: with this pattern, microservices not only gain in autonomy but also avoid direct coupling to shared databases.
The original Single Responsibility Principle (SRP) is about having cohesive functionality in an OO class. Having multiple responsibilities in a class naturally leads to tight coupling, and results in fragile designs that are hard to evolve and can break in unexpected ways upon changing. The idea is simple, but as Uncle Bob pointed out, SRP is very easy to understand, but difficult to get right.
The notion of single responsibility can be extended to the cohesiveness of services within a microservice. The microservice architecture style dictates that the deployment unit should contain only one service or just a few cohesive services. If a microservice is packed with responsibilities, that is, too many not quite cohesive services, then it might bear the pains of a monolith. A bloated microservice becomes harder to evolve in terms of functionality and the technology stack. Also, continuous delivery becomes burdensome with many developers working on several moving parts that go in the same deployment unit.
On the other hand, if microservices are too fine-grained, more likely several of them will need to interact to fulfill a user request. In the worst-case scenario, data changes might be spread across different microservices, possibly creating a distributed transaction scenario.
An important aspect of maturity in microservice design is the ability to create microservices that are not too coarse- or too fine-grained. Here the solution is not in any tool or technology, but rather on proper domain modeling. Modeling the backend services and defining microservice boundaries for them can be done in many ways. An approach that has become popular in the industry to drive the scope of microservices is to follow Domain-Driven Design (DDD) precepts. In brief:
- A service (e.g., a REST service) can have the scope of a DDD aggregate.
- A microservice can have the scope of a DDD bounded context. Services within that microservice will correspond to the aggregates within that bounded context.
- For inter-microservice communication, we can use: domain events when asynchronous messaging fits the requirements; API calls using some form of an anti-corruption layer when a request-response connector is more appropriate; or data replication with eventual consistency when a microservice needs a substantial amount of data from the other BC readily available.
IDEALS are the core design principles to be followed in most typical microservice designs. However, following the IDEALS is not a magic potion or spell that will make our microservice design successful. As always, we need to have a good understanding of the quality attribute requirements and make design decisions aware of their tradeoffs. Moreover, we should learn about the design patterns and architecture tactics that can be employed to help realize the design principles. We should also have a good grasp of the technology choices available.
I have employed IDEALS in designing, implementing, and deploying microservices for several years now. In design workshops and talks, I have discussed these core principles and the many strategies behind each with a few hundred software developers from different organizations. I know at times it feels like there is a landslide of tools, frameworks, platforms, and patterns for microservices. I believe a good understanding of microservice IDEALS will help you navigate the technology space with more clarity.
Many thanks to Joe Yoder for helping to evolve these ideas into IDEALS.
About the Author
Paulo Merson has been programming in the small and programming in the large for over 30 years. He is a developer at the Brazilian Federal Court of Accounts. He is also a Visiting Scientist with the Software Engineering Institute (SEI), a certified instructor for Arcitura, and a faculty member of the master program in Applied Computing at the University of Brasilia (UnB). Paulo often delivers professional training to fellow developers in the US, Latin America, and Europe. He is co-author of Documenting Software Architectures: Views and Beyond, 2nd edition. Paulo holds a BSc in CS from UnB and a Master of Software Engineering from Carnegie Mellon University.