- Defining the problem scope.
- Targeting the most applicable solution domain for the scope of the problem. Scope encompasses all extremes of the problem space, from the unlikely and rare scenarios to the very common scenarios, these extremes moderate the demands for resources, disk, processor and bandwidth. The art of good design lies in knowing how to tailor your solutions to the most applicable solution domain for the problem at hand.
- Implementing the solution for the applicable domain of importance that is most efficient as opposed to most expedient.
- If multiple solution domains must be covered ensuring seamless transition of solutions from one algorithm to the next.
The first and second points are most important, as you won't know how best to solve a problem if you can't define its' extents. Many times developers are unable to put their fingers on all aspects to a problem, this is unfortunate as it may severely restrict the solution they engineer as their ignorance of aspects of the problem that can be exploited for symmetry leads them to make inferior algorithm choices for the solution domain. How do you determine a problems extents? You make sure you test it at the extremes of performance using an old tool, a thought experiment. Consider a very unlikely high load or activity situation and then roughly define an algorithm that solves it (one solution domain), then consider the opposite , very low load and determine the optimum solution domain for that problem regime, finally determine a middle of the road situation and define a solution for that..depending on the problem, you may find that a given solution is optimal across the entire problem space or it may indeed require separate optimizations within the problem scope. Once you are done testing these edges or the event horizon of the problem, you have covered all the realizable conditions and thus can be assured you are engineering optimum solutions even if the eventual algorithm for most cases will not extend into the extreme scenarios discovered. In fact the act of defining the problem already sets you on the road to the solution, as by this task you also determine which of the identified solution domains are the most likely use case for the load, resource and bandwidth constraints of the final implementation.
The next two points cover the implementation which I like to call popcorn, the solution domains have been isolated, the optimal domain(s) for the problem in question explored using the previous intellectual muscle work and now it is time to just do the grunt work of building it. Now the mind shifts from looking at big picture concerns of latency between servers in a cluster to little picture concerns local to the executing machine during run time. A good example is in noticing how the choice of a variable declaration as static can effect memory utilization on a machine, other similar concerns are the choice of implementing a method as a concrete or as a forced overridden method from a abstract base class or interface inheritance. These choices can hinder or help the execution of code efficiently on any given system. One that I tend to pay particular attention to is byte size, every character in your code that is not needed is memory taken during execution, under loaded conditions these bytes add up to significant performance reduction so making your code as tight as possible through extreme parsimony of characters directly benefits efficiency in the long run. The rule of thumb I use is , use as many characters as required to ensure intelligibility of the code and no more. Another major source of issues lies in making classes too big, a class should only contain methods that are inately associated with the class. It makes sense for a "File" class to have a "read" method but it probably doesn't make sense for it to have its own "copy to" (as "copy to" should be something you do to Files not that Files do to themselves, a very subtle distinction), also note when a function that you wish to add to a class could also be useful to other classes. These generalized functions are better off in a static Utilities class where they can be employed on the different classes that need them, and where the static nature of the class ensures minimal loading of the method code for each instance of classes that employ the function. For example, if "copy to" was implemented in "File" it would be loaded to memory every time a File was instanced, taking up critical resources to do it, under load this seemingly small difference could prematurely curtail performance and directly impact operating cost. Moreover, loading the class instance (and all its methods) does not guarantee they will be used...so the loading is waste for most cases (especially for a File class where you most likely want to read from it, write to it rather than copy it to some location) By having the "copy to" method in a static Utility you ensure that it is highly likely to be used over the landscape and lifecycle of ALL classes that are in your class hierarchy that may require the function.