Martin Fowler says: “Don't even consider
Microservice unless you have a system that's too complex to manage as a
monolith.” The reason for his caution is that Microservice is not a
free lunch (nothing in life that is worthy is free): there is a great overhead
in implementing Microservice.
He shows the overhead in this diagram:
So where does the overhead lie? I like this diagram from http://microservices.io/patterns/index.html,
which shows that the Miroservice architecture involves much more than just application
design:
Underneath the application layer, there are the application
infrastructure layer and the infrastructure layer. In today’s technology, the application
infrastructure layer usually involves some highly robust message broker,
logging gathering and analysis tools etc; while the infrastructure layer
usually involves some service registry and discovery, docker orchestrating
tools etc; on the application layer, you have to consider how all the services
can behave nicely together, you have to consider API compatibility, failure
handling, eventual consistency etc.
So not only Microservice is not a free lunch, it is quite
expensive!
Martin Fowler is right to caution you about Microservice, but
I think he is wrong in saying that you should start with monolith. There are
great benefits in starting with Microservice in mind – even if you are building
a monolithic application.
I got this awakening when I tried to dissect one of the systems
I have worked on. I first listed its features, and picked one to practice. I thought
it wouldn’t be too hard because it has a very clear functionary boundary, but pretty
soon I found out it has a lot hidden dependencies with other functions of the
system.
In an aging system, hidden dependencies and relationships
are hard to find. We thought we were building a clearly layered beautiful system:
We thought we have walls along each layer and around each
component. But in reality, because everything is in the same code base and in the
same DB, it is so easy to break the walls and even tunnel underneath the walls.
Sometimes you do this without even a second thought especially when you saw
there were already holes in the wall. You ended up with this:
This is very bad: business logic is scattered in the code
and in the database models.
I wanted to break out moduleA as a Microservice, and
discovered its hidden dependencies through discussing with my colleagues and
digging into the database models. In the process, often times, discussions like
the following sprung up:
“Why did we design this model like this?”
“Probably to provide the maximum flexibility.”
“Do we have any customers using this function in this way?”
“Probably not many, but maybe a few…”
“So we went to so much trouble just for maybe a few…”
If we had started with Microservice in mind, we would have setup
walls around the bounded context. We might be building a monolith, but
comparing with a monolith without having Microservice in mind, the walls will
be guarded more carefully because we have the idea or hope that one day we
might break things apart into Microservices.
The conversation, in this situation, might be like this:
“We need this feature.”
“This feature requires communicating across walls, it is
expensive. Do you really need it?”
“I didn’t know it was so expensive. We need to think about this
carefully.”
Having visible walls will force us to be economic, it will
force us to weigh benefits and costs. It doesn’t mean that we can’t open holes
or create tunnels underneath to work around the walls – even in a Microservice
architecture, we can still make a mess, at the end of the day, it is our skills
that count. But having Microservice in mind force us to do deliberate thinking
and make decisions explicit.
The moduleA in my system has the following data models:
It has this structure because in many of its tables, either “flexibility”
is reserved for some unknown future usage or because other functions tried to
piggyback on those tables (e.g. adding some fields into the tables, the new fields belong to different domain models than the tables). Applying DDD, moduleA’s core can be structured as:
Of course, I am benefited
by 100% hindsight. But I believe if there is one single bullet, it is: start
with simple, evolve it into complexity after it is proven.
No comments:
Post a Comment