Saturday, February 21, 2015


I've been meaning to write up some thoughts I've had on microservices for quite a while but never really had the time (or made the time). However, when I was asked if I'd like to do an interview on the subject for InfoQ I figured it would be a forcing function. I think the interview was great and hopefully it helped to make it clear where I stand on the topic: whatever it's called, whether it's services-based, SOA, or microservices, the core principles have remained the same throughout the last 4 decades. What worries me about microservices, though, is the headlong rush by some to adopt something without understanding fully those principles. Yes there are some really interesting technologies around today that make it easier to develop good SOA implementations, such as Docker, Vert.x and Fabric8, but building and deploying individual services is the least of your worries when you embark on the service/SOA road.

Of course identifying the services you need is an important piece of any service-oriented architecture. How "big" they should be is implicitly part of that problem, though most successful implementations I've come across over the years rarely consider size in terms of lines of code (LOC) but rather business function(s). The emphasis on size in terms of LOC is one of the things I don't really like in microservices - unlike in the scientific world, micro here is not well defined and one persons "micro" could easily be another's "macro".

This brings us to the concept of service composition. Not all applications will be built from atomic (indivisible) services. Some services may be composed of other services. Of course there are different ways of approaching that from an implememtation perspective. For instance, a composite service could just be an intelligent endpoint (service) which acts as a coordinator for remote interactions on the other (constituent) services, such that a client never gets to interact with them directly. Another implementation could be to co-locate the consituent services in the same address space as the coordinator. Of course a good service-based implementation would not allow these implementation choices to be exposed by the composite service API to the client so that implementations could be changed without requiring changes to users.

Some other factors that must be taken into account when considering the "size" of a service or how the constituent services relate to a composite service include:

(i) performance - remote invocations, whether using a binary protocol such as AMQP, and especially text based protocols such as HTTP, are significantly slower than intra-process communication. Therefore, if performance is important and there will be many cross-service interactions, it may make sense to either consider merging services or deploying them within the same address space so that intra-process communication can be used instead. A service does not have to reside in its own distinct operating system process.

(ii) fault tolerance - despite the fact that you may be able to replicate a specific service to try to obtain high availability, there's always going to be a finite probability that your service will become unavailable - catastrophic failures do happen (entropy always increases!) And remember that it may not be possible, or at least easy, to replica some services, e.g., active replication of a service requires the business logic to be deterministic otherwise you need to use a passive replication protocol, which may adversely impact performance to the point of making it unrealistic to replicate in the first place. Therefore, if the failure of a service causes other services to be unusable (no Plan B), it may make more sense to co-locate these services into a unit of failure, such that if one is going to fail (crash) they all fail anyway.

Of course there are a number of other considerations when building applications from services and again we've been addressing them over the years in SOA deployments or earlier with services-based applications. One of the most critical is service (or task) orchestration - rarely is an application constructed from just one service and even if it were, that service may itself be using other services to perform its work. As such, applications are really the flow of control between different services and rarely is this some static a priori determination; the failure of a service may cause the application logic to determine that the next action is to invoke an operation on some other (compensation?) service. But again, as I said earlier, this is something we really should be taking for granted as understood, at least at a fundamental level. Whether or not the microservices community likes standards such as BPMN, the aims behind them or their bespoke implementations, remain and we would be ill advised to ignore. At least if you're going to build a new wheel it's a good idea to understand why the current wheel isn't good enough!

For a long time I wrote about and considered the necessity for SOA governance. In fact even before the term SOA was coined, governance has been a crucial component of any distributed system. Unfortunately it's also one of the most overlooked aspects. As we moved to a more service-oriented approach, runtime and design time governance became more and more important. How do you know the service you're about to use actually offers the right business capabilities as well as the right non-functional capabilities, e.g., can respond within the desired time, is secure, offers transactions etc? Of course there are a number of ways in which these questions can be answered, but essentially you need a contract between the client and the service. Part of the contract will inevitably have to include the service API, whether defined in WSDL, WADL, IDL or something else entirely. These days that part of (SOA) governance is not subsumed within the new term API Management. No less important, just a different categorisation. And microservices needs exactly the same thing because it really doesn't matter the size of a service - it'll have an API and hence need to be managed or governed.

Despite what I've heard about microservices, I really do believe that the existing SOA community had a lot to offer their microservices cousins; we've got to build on the experiences of the past decades and deliver better solutions to our communities rather than think that starting from scratch is the right approach.

No comments: