I am a tad frustrated about the whole µService discussion. How carefully designed concepts and terminology gets trampled over by a crowd that doesn't bother to look around in their rush to next year's fashion. When Neil Bartlett posted a link to an interesting blog that discusses exactly this subject I was therefore pleasantly surprised. It is about how our industry tends to swing from hype to hype and does not even try to learn any lessons from the down-swing. That is, if those lessons requires more than a minimal attention span. Ok, Simon Brown does not appear to be an OSGi aficionado but he hits the nail on the head that your monolithic big ball of mud is not going to be a tender agile application overnight by adopting micro-web-services. Or, how a new fashion is always more attractive than the grunt work of making the existing stuff right. Hey, maybe they did find the silver bullet this time! #not
We actually know quite well, and for a long time, how to build large complex systems: contract based component designs scale very well. It just so happens that in practice the siren song of short cuts are much more attractive than doing it right ... that is, in the short run. We show beautiful layered architectures that are in reality camouflage nets over the actual big ball of muds. Read Simon's blog and weep.
Now, although I agree with the view in this blog, I think he misses the key reason why we always end up with that big ball of mud. Unfortunately, experience shows that this is not an easy story that fits the 140 character attention span, the cause of the big ball of mud does require some work to understand. Please shut off twitter and bear with me.
In the seventies, structured programming was the then current best practice. The mantra's in those times were 'Low Coupling' and 'High Cohesion'. Then, in the early eighties, we got Objects! New! Improved! No need to bother about these old geezers with their coupling and cohesion muck anymore. (I never really got that cohesion stuff anyway.) Progress, New! We were young and innocent until time taught us that the old geezers had probably learned a lessons or two that we had missed. Embarrassingly, the anticipated reuse we had promised to our managers were much harder to achieve than we had thought in our youthful exuberance. What killed us most were the transitive dependencies that are implicit in the object programming models. It is wonderful to reuse another object, but the damned thing should then not do the same thing and depend on others, ad nauseum. Instead of picking and choosing reusable objects, systems quickly became big balls of mud, or more appropriately, a different form of spaghetti coding. We had carefully avoided the GOTO statement but the result was painfully similar, just on another level.
In the nineties I had learned my lessons about coupling, cohesion, and objects. So when I met Java in 1996 I immediately realized what interfaces were doing, they were addressing the transitive dependency problem that had caused me so much pain! With an interface you could program against a contract, not against an implementation! This did wonders for the transitive dependency problem since you were now bound to an interface. Interfaces were much less coupled by design than implementation classes. I obviously wasn't alone and interfaces became highly popular in Java and are pervasive today because they do provide freedoms that transitive dependencies kill.
Alas, we then witnessed a demonstration of how little intra-generational learning is going on. Two things went haywire. First, interfaces raised this pesky problem of 'How on earth do we get the instance in runtime?` Factories were kind of awkward so we found the greatest ostrich solution of all software times: 'We hide it in a XML!'. Obviously, the fact your coupling Java analysis tools did not see these hidden dependency anymore did not mean they stopped hurting you over time. An implementation class reference is a transitive dependency as much as any "new" expression is in your compiled code. Worse, a class referenced in text bypasses all those type safe-guards in Java you just spent all that money on.
The other, and surprisingly even worse, mistake was that we happily copied the transitive implementation dependencies model that failed so spectacularly with objects into our modules. And surprise, we got the same ball of mud, just now on a tad larger scale. Don't get me wrong, it is not maven that downloads the Internet, it is people that download the internet. People are addicted to solutions that look easy. That is, until that proverbial stuff hits the fan but then the original designers have then usually moved on.
Hard coding a dependency on another module will inevitably create more and bigger balls of mud, we learned this with structured programming, with object oriented programming, and now we are learning it with modular programming. The structural problem is identical in all these cases, so why should the result differ? (Though I find that a surprisingly large group of people disconnect now, 'What on earth has an interface to do with modules? This guy is an idiot ...'. Then again, their attention span probably already had made me loose them to that interesting tweet that just came in.)
Obviously, when we started designing OSGi in 1998 we did not have the clear view of today. However, the crew that did the early design had a lot of experience and provided a very elegant, µService based, model. It solved the 'how on earth do I get my instance of that interface' problem in a non-ostrich way, it also innovatively used Java packages as the dependency model. The Java package plays the same role for a module as the interface does for a class.
If you find this hard to swallow then you're not alone (which does not make you right btw, just more common). At JavaOne, I actually once overheard a discussion between three (at that time) Sun/Oracle high level executives; they were making fun of package dependencies, radiating an amazing ignorance of the underlying issues. Even with very bright people, some that I admire a lot, I've seen that it takes effort to accept this structural analogy between classes and modules. However, even at the JCP, the Java package is already playing a large organizing role in most, if not all, specifications. It acts as the mediator between the provider (e.g. Jersey) of the specification (JAX-RS in javax.ws.rs. packages) and the consumer of that specification.
The key consequence of the OSGi model is that the µservice is not an implementation (a module) as many people tend to talk about it, it is a contract. Modules can provide implementations for these service contracts and they can consume objects that implement these service contracts. The modularity model is thus not: module A depends on module B. No, module A depends on service S and module B and module C can provide such a service. It is this crucial indirection that I think Simon misses in his blog that is the only known answer against drowning in the big ball of mud over time. We figured this out with Java interfaces, we now only need to understand how we can eradicate the same transitive pest with modules as well.
Yes, I know the model is not as simplistic as the module A depends on module B model that we know so well; it adds an additional indirection that is not free. The difference is that you pay that cost up-front while the long term cost of module dependencies is much, much higher but drawn out over time. Though we like to postpone costs, we actually did learn a lot of lessons over time that made us take some up-front cost, Java as a type safe language is the prime example. So I am utterly convinced that model we learned in OSGi will prevail over time. Just hope that I do not have live through another hype that presents (parts of) these ideas as if they are something new.