Micro-services are one of today's leading software fashions; see this for more on software fashions in general. I have treated in depth the false claims of micro-services to make applications "scalable," and separately treated in depth the false claim that micro-services enhance programmer productivity.
Where the heck did such a crappy idea that has taken such a large mind-share in the software and management community come from?
I don't know. Who knows why skirts are short or long this year? Or whether shirts are tucked or not tucked? What I do know is that bogus fashion trends infect the software industry, some of them delivering decades of perniciousness. Looking at the spectrum of software fashions, I've noticed similarities among some of them. Some fashions well up, spread and peter out, only to recur again later with minor changes and a new name. The new name appears to be important, so that everyone is fooled into thinking the hot fashion is truly "new," not tarnished by the abject failures of its prior incarnations. I describe a clear example of an old idea with new name here.
The History of Micro-services
Why do we have micro-services? How did this bad idea get to be so popular? Was it someone's recent "great idea?" There are probably people who think it's their idea.
Do micro-services have a history? Of course they do! Nearly every major practice in software has a history of some kind. The number of software practices that are newly invented is truly tiny and going down rapidly. So why do we hear about software advances and the "new thing" so often? The whole field of software ignores its own history in truly radical ways. It even ignores what other software people in other fields and domains are doing.
Nonetheless, anyone who knows a broad range of software practices over a period of decades can easily recognize similarities and patterns — patterns that surely constitute "history" whether or not the actors recognize the precedents of what they do.
What are Services?
Let's step back and understand what "micro-services" are. By the name, it's obvious that a micro-service is an itty-bitty "service." So what's a service? Once you peel back all the layers, a service is a plain old subroutine call with a bunch of fancy, time-consuming stuff stuck in between the code that calls and the code that is called. See this for my explanation of the fundamental concept of the subroutine.
Subroutines (or functions or methods or whatever) are an essential aspect of programming. Every programming language has its syntax, statements and other things that people get all wrapped up in. Every programming language also has an associated library — as in function/subroutine library. If you look at a manual for a language and one for the associated library, the library is nearly always larger. The richness of a library is a key factor in the utility of a language. In modern terms, an associated framework serves essentially the same purpose; an example is the RAILS framework for Ruby, without which Ruby would just be one line on the ever-growing list of mostly-forgotten languages.
When you're writing a program, knowing and making great use of the associated subroutine library is essential. But what if you see that you're doing the same kind of thing over and over in your program? It makes sense to create a subroutine to perform that common function as part of the overall application you're building. As time goes on, most well-structured applications include a large portion of nested subroutines.
What if there's a program that can only run on another computer and you want to call it as a subroutine — give it some data to work on and get the results back? This problem was addressed and solved in many variations decades ago. It's most often called an RPC — Remote Procedure Call. To make a call to a distant subroutine you implement a version of it locally that looks and acts the same, but does some networking to send the fact of the call and the parameters to a cooperating routine on the distant computer. That routine gets the network call, retrieves the data, and makes a local call to the subroutine. When the return happens, the data is sent back by the same mechanism. It's just like writing a memo and putting in your out-box. The clerk who picks up the outgoing memos delivers the ones that are local, and puts the ones that aren't into an envelope, addresses it and puts the memo with return instructions into the mail. And so on. The calling code and the called code act like they normally do and the RPC makes it happen.
Along comes the internet. People notice that http is supported all over the place. Why not leverage it to implement the mechanics of the RPC? Voila! Now we have SOAP — Simple Object Access Protocol, which uses http and XML to send and return the data. A subroutine with a SOAP interface was called a "service." This all got standardized roughly 20 years ago. Then along came a simpler version to use that had even more overhead called Restful, which is still in widespread use.
If you need to call a subroutine that can't or shouldn't run on your computer but on some other computer, having some form of RPC is extremely useful. The time penalty can easily amount to a factor of thousands or even millions compared to a normal local subroutine call, but you'll gladly pay the price if there's no other way.
Now let's recall what all this amounts to. A "service" is a subroutine that has a HUGE amount of overhead. There is no reason to ever take on the overhead unless, in the vast majority of cases, the service you want to call is on a different computer than the one doing the calling, a computer that can only be accessed by some kind of networking.
Prior Incarnations of Service Insanity
There have been a couple waves of service fashion mania. One of the waves took place roughly twenty years ago during the internet bubble. People were very concerned to be able to support growth in the use of their applications by unprecedented factors. Expert opinion rapidly agreed that building a "distributed application" was the way to accomplish this. The idea was that the load on the computer would greatly exceed the capacity of even the largest, most capable computer to handle it. To anticipate this, the application should be architected to be able to run on multiple computers that could be changed at will. Instead of a "central" application, the application would be "distributed." This was further elaborated by replacing the simple RPC mechanism with queues called "service buses," as in a bus to route large numbers of service calls from one place to another. The bus itself had to be robust so that messages weren't dropped, so it evolved into an "enterprise service bus," something which exists today. A huge, complex mechanism for accomplishing this became part of java, J2EE (Java 2 Enterprise Edition). See this for the complex, reincarnating history of the transaction monitor.
It turned out that building a "distributed application" was vastly more expensive and time-consuming than just building a plain old application, the kind that today is sneeringly dismissed as being "monolithic." The computational and elapsed time running cost of such an application was also huge. Hardly any applications had loads so large that they needed to be distributed. Even worse, the few applications that ended up with huge loads that required many computers found vastly simpler, more effective ways to get the job done. It's worth noting that articles with titles like "Why were we all fooled into thinking building distributed applications was a good idea" never appeared. The whole subject simply faded away. Here's a description from another angle of the distributed computing mania.
Another incarnation of this nutty idea took place largely inside large enterprises. Many of these places had highly diverse collections of applications running the business, accumulated by acquisition, multiple departments building applications for themselves, etc. Instead of doing the sensible thing of reducing the total number of applications by evolving the best versions to be able to handle more diverse tasks, the idea of a SOA, service-oriented architecture, became the focus. All these applications could be turned into services! Instead of building new applications, people would now build services. Tools were built to manage all these new services. There were directories. There was tracking and control. All sorts of new demands came latching onto the IT budget.
SOA never got to the fever pitch of distributed applications, but it was big. It went the same way — fading as it turned out to be a lot of time and trouble with little benefit.
Now we come to the present. Remember that the original motivation of the RPC in any form is that a subroutine you want to call is ONLY running on some other computer. An RPC, whatever the form or cost, is better than nothing. Sensible then and sensible now. Then came ways of building programs so that their parts COULD BE run on different computers, if the computational requirements became huge. This also turned out to be rarely needed, and when it was needed, there were always better, faster, cheaper ways than a service approach. Now, with micro-services, we have reincarnated these proven-to-be-bad ideas, claiming that not only will micro-services effortlessly yield that wonderful virtue "scalability," but it will even make programmers more productive. Wrong and wrong.
Conclusion
I like the phrase "oldie but goodie." Sometimes it's applicable. In computing when a hot new tech trend comes along that "everyone" decides is the smart way to go, it rarely is. It most often is an "oldie but baddie" with a new name and new image. You would think that software people were so facts-oriented that they wouldn't be subject to this kind of emotionally-driven, be-part-of-the-group, get-with-the-program-man kind of thinking. But most people want to belong and want increased status. If believing in micro-services is the ticket to membership in the ranks of the software elite, a shocking (to me) number of people are all in.
