Defining methods via prototype vs using this in the constructor - really a performance difference?

Question

In JavaScript, we have two ways of making a \"class\" and giving it public functions.

Method 1:

function MyClass() {
    var privateInstanceVariable         


        

Answer 1

This answer should be considered an expansion of the rest of the answers filling in missing points. Both personal experience and benchmarks are incorporated.

As far as my experience goes, I use constructors to literally construct my objects religiously, whether methods are private or not. The main reason being that when I started that was the easiest immediate approach to me so it's not a special preference. It might have been as simple as that I like visible encapsulation and prototypes are a bit disembodied. My private methods will be assigned as variables in the scope as well. Although this is my habit and keeps things nicely self contained, it's not always the best habit and I do sometimes hit walls. Apart from wacky scenarios with highly dynamic self assembling according to configuration objects and code layout it tends to be the weaker approach in my opinion particularly if performance is a concern. Knowing that the internals are private is useful but you can achieve that via other means with the right discipline. Unless performance is a serious consideration, use whatever works best otherwise for the task at hand.

1. Using prototype inheritance and a convention to mark items as private does make debugging easier as you can then traverse the object graph easily from the console or debugger. On the other hand, such a convention makes obfuscation somewhat harder and makes it easier for others to bolt on their own scripts onto your site. This is one of the reasons the private scope approach gained popularity. It's not true security but instead adds resistance. Unfortunately a lot of people do still think it's a genuinely way to program secure JavaScript. Since debuggers have gotten really good, code obfuscation takes its place. If you're looking for security flaws where too much is on the client, it's a design pattern your might want to look out for.
2. A convention allows you to have protected properties with little fuss. That can be a blessing and a curse. It does ease some inheritance issues as it is less restrictive. You still do have the risk of collision or increased cognitive load in considering where else a property might be accessed. Self assembling objects let you do some strange things where you can get around a number of inheritance problems but they can be unconventional. My modules tend to have a rich inner structure where things don't get pulled out until the functionality is needed elsewhere (shared) or exposed unless needed externally. The constructor pattern tends to lead to creating self contained sophisticated modules more so than simply piecemeal objects. If you want that then it's fine. Otherwise if you want a more traditional OOP structure and layout then I would probably suggest regulating access by convention. In my usage scenarios complex OOP isn't often justified and modules do the trick.
3. All of the tests here are minimal. In real world usage it is likely that modules will be more complex making the hit a lot greater than tests here will indicate. It's quite common to have a private variable with multiple methods working on it and each of those methods will add more overhead on initialisation that you wont get with prototype inheritance. In most cases is doesn't matter because only a few instances of such objects float around although cumulatively it might add up.
4. There is an assumption that prototype methods are slower to call because of prototype lookup. It's not an unfair assumption, I made the same myself until I tested it. In reality it's complex and some tests suggest that aspect is trivial. Between, prototype.m = f, this.m = f and this.m = function... the latter performs significantly better than the first two which perform around the same. If the prototype lookup alone were a significant issue then the last two functions instead would out perform the first significantly. Instead something else strange is going on at least where Canary is concerned. It's possible functions are optimised according to what they are members of. A multitude of performance considerations come into play. You also have differences for parameter access and variable access.
5. Memory Capacity. It's not well discussed here. An assumption you can make up front that's likely to be true is that prototype inheritance will usually be far more memory efficient and according to my tests it is in general. When you build up your object in your constructor you can assume that each object will probably have its own instance of each function rather than shared, a larger property map for its own personal properties and likely some overhead to keep the constructor scope open as well. Functions that operate on the private scope are extremely and disproportionately demanding of memory. I find that in a lot of scenarios the proportionate difference in memory will be much more significant than the proportionate difference in CPU cycles.
6. Memory Graph. You also can jam up the engine making GC more expensive. Profilers do tend to show time spent in GC these days. It's not only a problem when it comes to allocating and freeing more. You'll also create a larger object graph to traverse and things like that so the GC consumes more cycles. If you create a million objects and then hardly touch them, depending on the engine it might turn out to have more of an ambient performance impact than you expected. I have proven that this does at least make the gc run for longer when objects are disposed of. That is there tends to be a correlation with memory used and the time it takes to GC. However there are cases where the time is the same regardless of the memory. This indicates that the graph makeup (layers of indirection, item count, etc) has more impact. That's not something that is always easy to predict.
7. Not many people use chained prototypes extensively, myself included I have to admit. Prototype chains can be expensive in theory. Someone will but I've not measured the cost. If you instead build your objects entirely in the constructor and then have a chain of inheritance as each constructor calls a parent constructor upon itself, in theory method access should be much faster. On the other hand you can accomplish the equivalent if it matters (such as flatten the prototypes down the ancestor chain) and you don't mind breaking things like hasOwnProperty, perhaps instanceof, etc if you really need it. In either case things start to get complex once you down this road when it comes to performance hacks. You'll probably end up doing things you shouldn't be doing.
8. Many people don't directly use either approach you've presented. Instead they make their own things using anonymous objects allowing method sharing any which way (mixins for example). There are a number of frameworks as well that implement their own strategies for organising modules and objects. These are heavily convention based custom approaches. For most people and for you your first challenge should be organisation rather than performance. This is often complicated in that Javascript gives many ways of achieving things versus languages or platforms with more explicit OOP/namespace/module support. When it comes to performance I would say instead to avoid major pitfalls first and foremost.
9. There's a new Symbol type that's supposed to work for private variables and methods. There are a number of ways to use this and it raises a host of questions related to performance and access. In my tests the performance of Symbols wasn't great compared to everything else but I never tested them thoroughly.

Disclaimers:

1. There are lots of discussions about performance and there isn't always a permanently correct answer for this as usage scenarios and engines change. Always profile but also always measure in more than one way as profiles aren't always accurate or reliable. Avoid significant effort into optimisation unless there's definitely a demonstrable problem.
2. It's probably better instead to include performance checks for sensitive areas in automated testing and to run when browsers update.
3. Remember sometimes battery life matters as well as perceptible performance. The slowest solution might turn out faster after running an optimising compiler on it (IE, a compiler might have a better idea of when restricted scope variables are accessed than properties marked as private by convention). Consider backend such as node.js. This can require better latency and throughput than you would often find on the browser. Most people wont need to worry about these things with something like validation for a registration form but the number of diverse scenarios where such things might matter is growing.
4. You have to be careful with memory allocation tracking tools in to persist the result. In some cases where I didn't return and persist the data it was optimised out entirely or the sample rate was not sufficient between instantiated/unreferenced, leaving me scratching my head as to how an array initialised and filled to a million registered as 3.4KiB in the allocation profile.
5. In the real world in most cases the only way to really optimise an application is to write it in the first place so you can measure it. There are dozens to hundreds of factors that can come into play if not thousands in any given scenario. Engines also do things that can lead to asymmetric or non-linear performance characteristics. If you define functions in a constructor, they might be arrow functions or traditional, each behaves differently in certain situations and I have no idea about the other function types. Classes also don't behave the same in terms as performance for prototyped constructors that should be equivalent. You need to be really careful with benchmarks as well. Prototyped classes can have deferred initialisation in various ways, especially if your prototyped your properties as well (advice, don't). This means that you can understate initialisation cost and overstate access/property mutation cost. I have also seen indications of progressive optimisation. In these cases I have filled a large array with instances of objects that are identical and as the number of instances increase the objects appear to be incrementally optimised for memory up to a point where the remainder is the same. It is also possible that those optimisations can also impact CPU performance significantly. These things are heavily dependent not merely on the code you write but what happens in runtime such as number of objects, variance between objects, etc.