In monolithic architectures, all features are implemented in a single codebase, with no visible modularization of the code. Objects reference each other in complex and intricate ways, making the code hard to read, understand and maintain:

Monolith Architect

modules or packages following the functional boundaries, with few dependencies (imports and method calls) between them. Some people call this a modular monolith :

Modular Monoli

Microservices are “just” one step further, splitting again each of the codebases into smaller parts: Finding the boundaries here is a bit trickier, as architects must deep-dive into the model to find (or create) meaningful seams. But we will talk about that later.

Microservices

Microservices allow for multiple scaling strategies, which can even be combined:<ul><li>vertical scaling: memory and CPUs are added to make things go faster</li><li>horizontal scaling: the load is shared between multiple identical instances behind a load balancer</li><li>sharding: the load is routed to a different instance depending on the data it needs to access</li><li>functional decomposition: the service is split into smaller services with technologies and scaling strategies specific for each kind of load</li></ul>

Features Of Microservices

Head of Labs and Delivery at Zenika. 
IT specialist focused on Software Craftsmanship, Agile and DevOps. 
Travel enthusiast and passionate photographer. 
Alumnus of ENS Paris-Saclay.

Tech conferences piling up presentations from large cloud-native companies, shiny new frameworks, consulting firms building up their portfolios, software engineers discussing how to ride the hype at the coffee machine… Microservices are everywhere. A lot has already been said and written about them, often making microservices look like a silver bullet. But why leave the comfort of the well-known, sometimes layered, sometimes modularized, monolithic architecture? Do you really need what microservices offer? Are you really ready to lose what microservices sacrifice? This article tries to summarize my view on the “microservices compromise”, what it requires and what it implies, beyond splitting a codebase into small pieces.

For the road to microservices is dark and full of terrors.

It is always good practice to define concepts before using them. So, what is a microservice? What are microservices architectures?

Consider the following functional architecture diagram of an e-commerce platform, resulting from the hard work of an architect somewhere.

Related features have already been grouped together, and dependencies between sets of features are illustrated with arrows.

In monolithic architectures, all features are implemented in a single codebase, with no visible modularization of the code. Objects reference each other in complex and intricate ways, making the code hard to read, understand and maintain:

Teams usually end up in this tough spot because of features piling up without a clear architectural strategy. Indeed, this situation is far less likely if the architecture diagram pictured above is already drawn and shared with the team!

In this case, we can hope for modules or packages following the functional boundaries, with few dependencies (imports and method calls) between them. Some people call this a modular monolith :

Going further on that path, the next logical approach is to split this large codebase into multiple parts. The module/package boundaries are a good place to start, but there might be some exceptions.

With this approach, people usually end up with what I call large services :

Microservices are “just” one step further, splitting again each of the codebases into smaller parts:

Finding the boundaries here is a bit trickier, as architects must deep-dive into the model to find (or create) meaningful seams. But we will talk about that later.

On paper, microservices architectures offer a lot of benefits…

It is known that small systems are easier to reason about and build. Who has not noticed that software is never produced more efficiently than during the first few weeks of development? This also holds for maintenance: regardless of team seniority and turnover, it is always easier to keep alive and change behavior in a small codebase than in a large plate of spaghetti code. Eventually, all software is retired, either because too many mistakes were made and the team decides to start over, or because the requirements or the environment changed too much. Again, replacing multiple small services is easier than redesigning an entire software, but with multiple independent pieces this work can also happen progressively, over the course of multiple months, and without blocking other development efforts. The times of “we need to stop adding new features to the product to refactor the codebase foundations in depth” are over.

Good news then! Microservices architectures are all about splitting software in multiple services and keeping them small. When done right, this code split follows the lines of the various concerns handled by the application. This means that each service is focused on a small, specific set of features, and does not have to care too much about the rest of the application logic. The benefits of such a decoupled architecture are obvious: ease of development and evolution, fewer dependencies between teams and roadmaps, fewer bugs and anomalies…

In a monolithic architecture, the choice of language and framework is made once, usually at the very beginning of the project. Librairies tend to accumulate as new requirements arrive or because team members change, and so do their preferences in tooling. This is why all monolithic architectures usually go for a large, generic framework (e.g., Spring for Java codebases) and a standard storage solution (e.g., a relational database).

“Jack of all trades, master of none” , as the saying goes. Despite all their (often impressive) efforts, how could such frameworks provide, at the same time, the optimal solution for use cases requiring a fast startup (async tasks), best performance during a few minutes or hours (batch jobs) and everything (from high frequency/low complexity to low frequency/high complexity) in the spectrum of modern web application workloads? Similarly, how could a database offer, at the same time, the best options to store structured, unstructured and binary data, regardless of response time, volume, scaling and consistency requirements?

The issue does not stop there, and soon enough a new, modern framework will appear that looks better suited for the job at hand. At this point, there will be no good option anymore: should you stick with the old, eventually abandoned framework, or rewrite the entire application? You will be stuck.

In microservices architecture, the rules are different. You can choose the right stack (language, frameworks, librairies, storage…), at the right moment, on a use case (microservice) basis. This could mean using, at the same time, a compiled language with no framework (e.g., Go) for services for which startup time is crucial, a lightweight web framework with an efficient, distributed cache (e.g. Node, Express and Redis) for high frequency/low complexity workloads, and an advanced framework in a typed, interpreted language (e.g., Java and Spring) for low frequency/high complexity use cases. The same logic applies to storage, and microservices also open the door to using, for instance, a columnar database alongside a relational database for analytics and transactional workloads.

Monoliths offer what I call “all-or-nothing” availability. Taking aside a few uncommon cases, the application is either up and running, or completely down. This situation is quite easy to reason about, since the resilience of the system boils down to your ability to keep the monolith process running. If anything goes wrong (host failure, crash, network failure, insufficient disk space…), the whole application is gone.

With microservices, a failure only affects one or a few of the multiple parts of the system, leaving most of the services running. This means that you now have the opportunity to react to partial failure and offer a degraded (but still useful) service to your users instead of a global error message.

Regarding scalability, monoliths can only grow vertically. That is, the only way to make a monolith faster or handle more requests is to make it run on a faster or bigger host.

Microservices, on the other hand, allow for multiple scaling strategies, which can even be combined:

<ul><li>vertical scaling: memory and CPUs are added to make things go faster </li><li>horizontal scaling: the load is shared between multiple identical instances behind a load balancer </li><li>sharding: the load is routed to a different instance depending on the data it needs to access </li><li>functional decomposition: the service is split into smaller services with technologies and scaling strategies specific for each kind of load </li> </ul>

Of course, all scaling strategies are not always available. For instance, horizontal scaling requires the service to be stateless, and sharding requires the computation to be localized enough for the underlying data to be sharded. One could also argue that some of these scaling strategies are also possible with monoliths, but to me this is already a step on the road to microservices!

In any case, in microservices architectures, the resilience and scaling strategies can be chosen independently for each service: a batch microservice could scale vertically and not have resilience requirements beyond a “retry again later”, while a web server is designed from the beginning to use sharding and scale horizontally. None of these options will ever be available in a monolithic architecture.

Finally, keep in mind that resilience and scalability decisions do not only pertain to services your team develops, as other components (databases, message brokers, API gateways…) come with their own tradeoff and scaling options, with a definitive impact on the behavior of the overall system.

As the size of a software grows, the size of its supporting team tends to grow accordingly, often with the incorrect assumption that more people implies more features delivered. It is well documented that the effort required to synchronize team members is not linear in the number of people, but rather quadratic: to synchronize 2 people, 1 phone call is enough; to synchronize 5 people, 20 one-to-one phone calls are required. Large meetings only solve this issue up to a point, arguably around 5 to 10 people.

Microservices offer the opportunity to naturally split the work between small teams, each focused and knowledgeable about their microservices and the few they interact with. Although collaboration and synchronization points between teams is still required, their frequency and duration should be dramatically reduced.

After reading the previous section, you will probably really want to use microservices in your current project and all the other ones, past and future! But as everybody knows, there is no silver bullet and no free lunch, only compromises. So let me sing you a song that is rarely sung, the one about the road to microservices, and everything you must be ready to lose along the way!

Models in monoliths are not created as a large pile of coupled objects by design. They progressively but inexorably get there, because concepts in the real world are inherently coupled, and simplifying reality to extract and maintain over time a simple, decoupled model is no easy task (it can actually be a full-time job). In monoliths, coupling can stay unnoticed for quite some time, before it eventually surfaces affecting development velocity, ease of maintenance or performance.

In microservices, there is nowhere to hide. With parts of the model distributed across the network, the cost of coupling is orders of magnitude higher, and will almost instantly result in maintenance burden and performance bottlenecks. A lot more effort will also be required to fix the model: instead of refactoring a co-localized set of objects, multiple services will now need to be merged, split and rearranged, with a global impact on the architecture.

Do not expect decoupling to come for free from splitting a large codebase into small pieces. In fact, it works the other way around: the split into small pieces will only work if your team already has the ability to design a decoupled model. If your monolithic codebase looks like a plate of spaghettis, it is therefore a strong hint that you lack the modeling skills required to succeed in microservices.

With microservices, the model now has boundaries that are hard to move, and the possibility of a large scale redesign is essentially lost. If you get these boundaries wrong, you end up with a lot of small, coupled services, which is absolutely worse than a large, coupled model in a monolith.

So here is the compromise: how confident are you in your ability to extract and maintain a useful, simple, decoupled model? And how much splitting into small pieces are you willing to bet on it?

Like in everything else, flexibility in technology options is great, but it comes with a price. Instead of one technology stack to master, you now have two, three, five, ten of them, each possibly made of multiple languages, frameworks, librairies and tools! And we are not only talking about development skills, but rather about a company-wide matter, from development and testing to integration, build tooling, deployment, configuration, monitoring, performance management, security… Having multiple technologies in production means having to maintain a pool of experts for each of them, in each department. Otherwise, the next rewrite from scratch will not be because of code quality but because of a lack of skills to implement simple changes.

Letting everyone use the languages and tools they like and hope for the best is clearly not a valid strategy, and some guidelines must be established. They can take the form of a reduced set of technology stacks to choose from (e.g., Netflix seems to focus on JVM technologies), or a particular process to test, validate and train people on new languages or frameworks.

Too many constraints, and you don’t get the advertised benefits of technology heterogeneity; too few, and the fun at the technology zoo will not last long! Quite obviously, this is again a matter of compromise.

Compared to monoliths and their all-or-nothing availability, microservices bring a different equation to the table. Because they are composed of more moving parts (multiple hosts, multiple services, multiple network links between them…), they are statistically more likely to fail. Pushed to a certain scale, a microservice architecture is, at any point in time, statistically more likely to have at least one failing component than to be entirely healthy. Seriously, do the math. Assume a 99% overall uptime for all your components:

<ul><li>with only one component, the probability to be healthy is, well… 99% </li><li>with N components, the probability to be healthy is 99%N </li><li>70 components are enough to have a &lt;50% chance to not have any failure at a given time! </li> </ul>

Microservices actually bring more failure, not less. But they also bring the opportunity to keep them localized and compensate for them. Again, it does not come for free, and there is extra work to be done to get there!

The first solution that should come to mind is the Bulkhead pattern, in which multiple, identical instances of a service are used behind a load balancer to handle traffic. If one of the instances crashes, the load balancer stops sending requests to it until it recovers (in some advanced versions of this pattern, the failing instance is automatically killed, and another one is spawned). Beyond the infrastructure and configuration work implied here, there is also a hidden requirement that the service handling traffic needs to be stateless. If this is not the case, you also need to invest time to get there. Of course, this is just for one pattern; other ones (for instance, Circuit breaker ), come with their own hypotheses about your system, and implementation burden.

In any case, unhandled errors have a greater impact in distributed systems, as you cannot simply rely on the nice transactional properties of a shared database. Without transactions, if one thing fails in one service, you have to explicitly undo something else in another service. And what if this undo action also fails? Some patterns, such as 2-Phase Commit or Sagas , can help here, but they also require extra work and come with their own trade-off, especially regarding coupling.

So, in the end, do you prefer investing in patterns to fight against the cold reality of statistics, or reducing the number of moving parts? Who said “compromise”?

Code is easier to change than organizations or culture. And cutting the code into small pieces will not magically change the way people work together. If your entire enterprise is organized around technical boundaries (e.g., a frontend team, a backend team, and an ops team), massive efforts will be required to evolve to a lot of small, focused, multidisciplinary teams. Imagine, 10 teams with each their architect, frontend specialists, backend developers, performance expert, ops engineer, designer, product manager… How much effort would be required to go from where you currently stand to this vision?

In a microservices setting, transversal teams are an anti-pattern, since they introduce, in their own way, coupling. I am not saying that you should do away with all the transversal teams, as the organization can choose to retain common guidelines (e.g., architecture, modeling, management…) or capitalize on some technologies (e.g., orchestration or monitoring tools). But to reduce the coupling, transversal teams should really act as communities of practices (rather than as a pool of resources to deliver features) and be located as close to the product teams as possible.

I will not go into the lengths of how hard change management is, but here goes for the compromise: how much energy into transforming the company’s culture, and how much into delivery features to real users?

At the end of the day, one should really wonder: what is there to win, and what am I ready to lose?

The only real guarantee of microservices architectures is to end up with a lot of small parts, lose transactional consistency, raise failure rates and challenge your organization.

Decoupling, resilience, scaling, tech heterogeneity and small focused teams will not come for free; they can only be the result of a tremendous company-wide effort, from the leadership to the infrastructure teams and back, regarding people, processes and tools.

On the one hand, only with all these options in hand will you be able to succeed in the challenges that come with modern visions and large-scale projects. But on the other hand, without an impressive amount of upfront work and change in your organization, your microservices project will almost surely fail.

With a coupled model, a single unified tech stack, no resilience or scaling strategies and only transversal teams, you will get the worst of distributed systems, for no actual benefits.

Therefore, before actually leaving the comfort of the monolith and jumping on the microservices bandwagon, be sure that you have what it takes to succeed. In fact, if you are currently failing to build and maintain a monolithic architecture, what makes you think that you will do better with microservices? It’s all the same, but way harder, in almost any imaginable way.

Microservices only show you a road that was previously hidden. Walking on that road is all on you, and it will not be easy.

So I hear that you know what you are doing, have quite some time and money in front of you, and decided to walk that road.

As you will see, this is no straight road. There will be many paths to explore, and many directions to choose from. This chapter is a short guide for some of the many crossroads you should prepare yourself to encounter.

For a variety of topics, mostly technical for now, we will indeed propose a spectrum of practices, tools and patterns, to compromise over. They will range from large, coupled, but simple approaches, closer to what we find in monoliths, to the small, decoupled but usually more complex ones, pertaining to microservices approaches.

As we deep dive into each topic and options, we should not lose focus on what we are trying to achieve: decoupling (D ), resilience (R ), scaling (S ), tech heterogeneity (H ) and small, autonomous, focused teams (T ). Even though it will sometimes look like an oversimplification, we will do our best to grade each practice or pattern depending on their alignment with these 5 objectives: either aligned (✔), misaligned (✘) or partially aligned, depending on implementation details or other factors (? ). This should give you a good first idea of where you stand, and where to go next.

How small the pieces we cut our monolith into?

What are the corresponding options for our domain model?

What constraints are imposed on the choice of technologies (languages, frameworks, persistence…)?

Do services share the persistence mechanism?

How do services communicate with each others?

How do the services react to the failure of one of their dependencies?

What data consistency guarantees does the overall system offer?

How is the project versioned, built and packaged?

In which format are the services packaged and deployed?

Are services deployed independently?

Is the frontend (UI) also split into modules or services? If so, with what granularity?

How is authentication implemented?

Bite-sized knowledge
to upgrade
your career

The road to microservices · Romain Vernoux

Monolith Architect

Modular Monoli

Microservices

Features Of Microservices

IDEAS ABOUT

MORE LIKE THIS

10 Interview Questions Every JavaScript Developer Should Know

12 Agile Skills for Software Developers

Can Dynamic Sites Go Serverless?

It's time to

READ

LIKE

A PRO!

Bite-sized knowledgeto upgradeyour career

The road to microservices · Romain Vernoux

Monolith Architect

Modular Monoli

Microservices

Features Of Microservices

IDEAS ABOUT

MORE LIKE THIS

10 Interview Questions Every JavaScript Developer Should Know

12 Agile Skills for Software Developers

Can Dynamic Sites Go Serverless?

It's time to

READ

LIKE

A PRO!

Bite-sized knowledge
to upgrade
your career