The Cost of Frameworks
aerotwist.com
aerotwist.com
Advocate something you don't do: hypocrite
Close.
Front-end frameworks are one solution to the problem of managing the complexity involved in making an app.
After working with a few of them, I'm not really sure that they're the best ones.
- Functions
- Abstraction
- Modularity
- Design patterns
- Programming paradigms
- DSLs
- Conventions for naming/formatting/arrangement
Is it my impression or is Paul on some sort of a crusade to downplay the dev ergonomics of React and try to convince people that it's "slow"?
TodoMVC benchmarks have been done before: https://github.com/pygy/todomvc-perf-comparison . So sure, there's room for performance improvements in mainstream frameworks, and React is not the fastest thing in the universe, but come on. Maintaining a large project in vanilla js is largely equivalent to writing code in assembler when there are good C compilers: it's doable, and required in a handful of situations, but not really wise for 99% of real world projects.
Re: Tom's response: Something that he didn't mention (which is not surprising since he's an Ember dev) is that frameworks do sometimes detract from end user experience by imposing "opinionated" complexity and assumptions that might prevent devs from doing certain specific things and settling with suboptimal UX - the old adage of "if you want to deviate from the holy way(tm) you're on your own"
I'm kinda in the middle of the two opinions: it's definitely important to have access to the "metal" (both in terms of actually being able to code against low level APIs, and in terms of the amount of effort required to wade through framework abstractions in order to get there), but even using vanilla js, a complex app does need a "framework" (in the sense of having rules for where things should be and how they should interact with one another, and in the sense that any non-trivial app will have "library-level" plumbing). So, why not meet in the middle and use a lightweight framework that does 95% of things well enough to actually be used in non-trivial mobile apps[1] but that doesn't have high enough byte count to be bloated?
So the real choice here is not vanilla vs framework but do you want to use someone's framework or build your own.
Yes using a JS Framework is going to be slower than writing and optimizing code by hand. In the same way that you can do faster code in assembler than using any languages or that using GTK on top of X is going to be slower than writing directly to the graphic card.
Yet we still uses them because it mean less bugs (because you are not reinventing the wheel every time), it encourage code reuse, it add structure, allow other dev to easily enter the code and open you to a vast library of modules. But most importantly it allow shipping features faster and that's where it impact the users.
Users care more about having an usable product than a fast one that does nothing (with some exceptions obviously). We can afford to optimize everything by hand and ship at the same time.
Sure, in some use-cases it will be too slow, but then you identify those particulars cases and you take your time to optimize them. Even by bypassing the framework if you need to.
But don't decide to not use a framework because the general use case will be a bit slower.
Or as Knuth put it decades ago "Premature optimization is the root of all evil"
The complete quote being, because it's relevant here:
> "We should forget about small efficiencies, say about 97% of the time: premature optimization is the root of all evil. Yet we should not pass up our opportunities in that critical 3%"
If it would take 25ms in vanilla JS (40x quicker) I'd still consider it a small efficiency gain in our case (from a users perspective).
---
(should mention that there is no network latency, since it's an installed app the JS-code is bundled, obviously)
And this coincided almost perfectly with the rise of the web and the end of desktop GUI supremacy. The few big desktop successes in the 2000s (uTorrent, DropBox) often used no framework at all or had minimalist UIs where they used little-known corners of the OS but moved the heavy interface lifting to the web.
I'm wondering if this is part of the general pattern of things reaching perfection just as they become obsolete. Right when a technology reaches mainstream adoption, everyone has an opinion about how to do it "right", and right as it reaches mainstream adoption, all the major opportunities have been plucked clean.
Our product, Elevate Web Builder, uses an in-memory element representation with a DOM change management architecture that avoids all of these issues. However, because of this design, you have to use our product and our framework in order to benefit from it, because it sits as a virtual layer on top of the DOM.
What the DOM really needs, but probably won't get because there could be some serious side-effects from bad coding, is a set of simple reference-counted beginUpdate/endUpdate methods on each DOM element that isolate that portion of the DOM tree from repaints, but not layouts. That way, dimensional information is always immediately available, but painting is always handled as a single last step.
I'm meandering here, but the point is that improvements in the DOM management layer of the browsers could drastically cut the necessity of frameworks whose primary purpose is to work around deficiencies in the DOM. Fix the browsers, and you fix the load time. Cutting out the frameworks just ends up trading different kinds of pain...
None of the major frameworks solve this problem. The closest may be React, which strongly discourages you from interacting with real DOM nodes or doing anything stateful, but even then, if you really want to touch that offsetWidth you can trivially destroy React's performance. Maybe the best alternative is just strict coding standards that say "Don't do that!"
I'd talked with a few Chrome PMs when I was working on Material Design for Google Search and discussed the possibility of some sort of batching API for layout-causing operations. They weren't against it in principle, but nobody could think of an API that was something we'd actually use as web developers. The problem is that thinking about your app in terms of "This expression causes layout, this expression dirties the DOM" is a big cognitive load, and making those two classes of expressions non-interleavable means you suddenly need to structure your whole app in different ways.
There'd also been talk of doing a subset of HTML that includes only the operations that can be done quickly, but to my knowledge, that never went anywhere. You run into problems with the whole huge installed base of the web; if you're going to throw everything out and start from scratch, why not just use a native mobile app, or even raw OpenGL ES commands?
What virtual DOM does is simplify application development by removing the concept of time, so the programmer doesn't have to think about how to mutate the existing state to get to the new state. It's the cheapness of creating new virtual nodes that's interesting [1], and quickly applying them to the DOM with diffing.
"The DOM already does dirty-checking on manipulation. Manipulating the DOM is cheap, it's manipulating the DOM while also getting dimension or style information that's expensive."
Yes, absolutely, and I wasn't clear on this point. The catch, however, is that once you start using JS for dynamic rendering of UI elements, as opposed to static HTML, this becomes quite the problem. Simple things like dragging an element around or other types of basic interactivity require knowing where things are in real time.
"None of the major frameworks solve this problem. The closest may be React, which strongly discourages you from interacting with real DOM nodes or doing anything stateful, but even then, if you really want to touch that offsetWidth you can trivially destroy React's performance. Maybe the best alternative is just strict coding standards that say "Don't do that!""
The only way I know of around this problem is what we did: maintain a virtual element for each actual DOM element, and have the virtual element act as a gatekeeper to the DOM. Want to read the width ? No problem, here it is as a simple integer property, and the DOM isn't touched. But, as you can imagine, this sends you down the road to hell pretty quickly, and we ended up managing everything manually. We effectively use the browser as a display device. And, of course, the code size is larger than we would like...and we're back to the topic at hand. :-)
"I'd talked with a few Chrome PMs when I was working on Material Design for Google Search and discussed the possibility of some sort of batching API for layout-causing operations. They weren't against it in principle, but nobody could think of an API that was something we'd actually use as web developers. The problem is that thinking about your app in terms of "This expression causes layout, this expression dirties the DOM" is a big cognitive load, and making those two classes of expressions non-interleavable means you suddenly need to structure your whole app in different ways."
Having come from many years doing desktop development on Windows, I've become accustomed to BeginUpdate/EndUpdate as it's a pretty standard way of batching updates that cause repaints. Of course, Windows also punts on layout, so there's a pretty big difference right there.
"There'd also been talk of doing a subset of HTML that includes only the operations that can be done quickly, but to my knowledge, that never went anywhere. You run into problems with the whole huge installed base of the web; if you're going to throw everything out and start from scratch, why not just use a native mobile app, or even raw OpenGL ES commands?"
I would say that you're leaving out a whole host of things that make the browser such an attractive host environment for applications, absent automatic layouts. There's the painting capabilities alone. Ever try to deal with transparency in Windows ? It's a nightmare, but in the browser it just works. The DOM element tree model itself is beautiful and a perfectly natural way of visualizing a UI. In general, it's just a more complete environment for the developer, and there's no arcane API to work around: access to the built-in functionality is just an object property away. Then there's the single codebase, easy distribution/installs.....
I just think that browser developers need to actively push towards making the UI functionality more generalized and display-oriented, and less specific to its roots of static documents that always require some sort of layout. If a piece of JS code places an absolutely-positioned element somewhere in the browser window, there should be zero penalty for reading its layout information after doing so.
fooify(element) {
element.offset().then((rect) {
// do stuff
otherelement.offset().then((rect) {
// do more stuff
});
});
}
fooify(element) async {
var h = await element.offsetHeight;
// do stuff
var h2 = await otherelement.offsetHeight;
// do more stuff
}
> There'd also been talk of doing a subset of HTML that includes only the operations that can be done quickly, but to my knowledge, that never went anywhere. You run into problems with the whole huge installed base of the web; if you're going to throw everything out and start from scratch, why not just use a native mobile app, or even raw OpenGL ES commands?
Interestingly, this was basically the origin of Flutter. It started as a fork of Blink with the worst things removed, but still used DOM and JS, but they kept removing and removing, eventually removing CSS and the DOM entirely, and replacing JavaScript with Dart. Now things are fast, but they aren't like the web either.
Are you in a latency-critical domain, which Google Search is? Then optimize to minimize the number of bytes shipped to the user and startup time. Do you have 100+ screenfuls of information? Then optimize so that making new screens is easy. Do users load your app once, leave it open in the background, but then need to perform a number of interactions quickly? Then ship down the initial interface quickly and progressively lazy-load bundles of functionality that are just a JS function away on click, like GMail does.
Incidentally, one of the main reasons Google Search finally adopted Closure was because it had to integrate with Google+, which was all done using Closure. So again, the choice of a framework was driven by product concerns.
It's not a failure when a company builds a home-grown framework that is carefully tailored to the product needs that they've discovered over the last few years. Indeed, having the resources and domain knowledge to build your own framework is a far bigger success than building an app with the hot framework of the month and failing because you look just like every other app out there.
Constructing your own framework has cost. Sure, you will acquire a lot of knowledge doing it, but is it going to help the bottom line? Maybe.
Once again, are you seriously advocating that _everyone_ should do "vanilla JS" without any frameworks?
Rewriting your product is not failure, particularly when done at the very early stages where it only takes a week or so anyway.
As a nice side benefit, this also avoids all the holy wars about which framework is best, since you can put off the decision until you have hard data about which is best for you.
1.) Just by choosing to use a framework, you incur costs: download time, initialization time, complexity & bug surface area, etc. You should make sure that you gain benefits commensurate with those costs. If you start by picking a random framework, then a.) how can you measure what the costs were? and b.) how do you know what sort of benefits would be most useful to you?
2.) Relatedly, starting from a baseline random framework can teach you lessons that are not true, and then you base your product decisions on that. For example, there's a common meme that you cannot have smooth 60fps animations on mobile websites. This is not true; however, getting smooth 60fps animations on the mobile web requires careful attention to which elements will render on the GPU and which won't, and no framework currently in existence will help you with that. Back in 2008 when I did my first startup, it was generally believed that Javascript was too slow for games; not true, JQuery was too slow for games. Doing things that others believe to be impossible is the essence of competitive advantage.
3.) Frameworks make certain tasks easy at the expense of making other tasks hard. If you start with a random framework, you will be incentivized to do things in the same way that everyone else who uses that framework is. That's why almost every web 2.0 startup started from around 2007-2010 looks like a frontend over a database. Looking like everyone else is a recipe for failure in business.
4.) The framework authors are all subject to the constraints of the browser; however, browser vendors (and people who program directly to web APIs) are not subject to the constraints of the framework. The set of programs that you can write efficiently without a framework is a strict superset of the set that you can write efficiently with one.
5.) It's easier to add code than to remove it. If you start with no framework and then decide that you need one, you've written only the code that you actually did need for your problem. If you start with a random framework and then decide it's not appropriate, you need to backtrack and identify what you were actually trying to accomplish before you shoehorned it into the framework.
2) going off 1), one could come to the (obviously ridiculous) idea that AJAX is really hard. Or that frontend build systems are not worth the trouble (I actually see this "myth" in the wild, and people wasting time due to it). 60fps on mobile is really not a "competitive advantage", even if you're going into the hyper-saturated gaming market (and in that case, you really ought to be writing ObjC/Java if you care about speed at all). Competitive advantage is about creativity and relentlessly exploiting opportunities, it's rarely ever the case where an obscure piece of software trivia will make a huge difference.
3) frameworks exist on a spectrum: if your argument was true, one could just pick an obscure framework. But honestly, It strikes me as wishful thinking to suggest that framework choice is the primary driving factor for a product (that's why the role of product managers exist)
4) there's a very real and recognized constraint that affects all low level systems compared to higher-level abstractions: complexity. jQuery is a DSL for avoiding 10000+ LOC DOM API code. Virtual DOM (or any "retained mode" templating system) help avoid 10000+ LOC jQuery spaghetti. Component systems let you create more DSLs, etc, etc.
5) it is easier to add code than to remove it, and that's precisely why most systems in the wild have technical debt: organic growth. When you start from scratch, in my experience (especially w/ a team over a period of time), it can get pretty hard to untangle the various subsystem out of an organically grown codebase. We had this issue in my previous company: there was a monolythic architecture for a mission critical system, and the 10 year old Session class was essentially impossible to replace without breaking everything. Speaking from experience, frameworks actually make it easier to bail out because there are only so many idioms you need to expect, and code is more-or-less organized even if it is a shitstorm by any other metric. With non-framework code, you get to spend a lot of time evaluating for the umpteenth time why exactly foo is seemingly replaceable, but actually not without breaking bar, baz and quux.
Frameworks can have their own complexity and developer overhead, they can demand you spend more time solving the problems they impose (cough AngularJS), they can be ill-suited to the requirements and/or problem domain you're working in.
Building your own isn't always the sensible choice, but it can be.
But there's one more consideration...isomorphic rendering. If a framework can be rendered on the server-side then the time-to-interaction only matters if it's longer than it takes for the user to mentally process the page and take an action.
Turbolinks should have been an extension, not in the core. With these kinds of gimmicks, and constant breaking of APIs from versions to versions , Rails is committing adoption suicide.
Yes, but time-to-interaction can be long when connectivity is poor. Connectivity is frequently poor.
My thinking is, if the frameworks do similar things, i.e. provide similar functionality, maybe it would make sense to have browsers provide (standard) APIs for (some of) that functionality.
Frameworks and libraries can then build on top of that to provide templating, data-binding, additional lifecycle stages, and other helpers, which Polymer and X-Tags do.
The problem is, software development is a moving target. The current feature-set of the web platform would let you easily build a state of the art application... in 2005. But things have moved on. People's ambitions and expectations for the functionality, responsiveness across devices, performance, aesthetic appeal, touch friendliness, accessibility and offline capability of their web apps has skyrocketed. And at the same time, we've learned more about the coding strategies and patterns that work (and don't work) for writing, large, ambitious web applications.
This is a good thing, because it means the web is moving forward, and so are we. However, it also means we're never going to reach a promised land where the web platform does everything we could ever need it to. There will always be new demands and new technology — retina screens, VR, fingerprint scanners, etc. — and the web platform will need to catch up in those areas. There will always be new frameworks, like React, that overturn existing best practices and experiment with new ways of doing things. This means will always be a need for libraries and frameworks at the cutting edge of web development, to pioneer the paths that can later be paved via standardisation.
We also need to be wary of premature standardisation. Web components have arguably suffered from this, although its proponents couldn't have predicted it when they began. For years, component-based encapsulation libraries/frameworks gained very little traction on the web, and so Google launched an effort to deliver components natively via a set of standardised APIs. However, halfway through their effort, Angular and React blew up, and suddenly everyone was writing components as directives and JSX components. This has caused some friction, because components as implemented by these frameworks (and thus coded by the majority of web devs) don't quite match the vision set out in web components. Web components will still deliver useful features, such as Shadow DOM, which will find use in these libraries, but had the web components effort started after Angular and React had appeared, its design would likely have looked somewhat different.
There are more frameworks in the world than Javascript frameworks, if you are only going to discuss a subset of them you should make this clear in the headline/introduction. A heading such as "The Cost of Javascript Frameworks" would be appropriate.
Years ago when browser compatibility required 100s of workarounds for each browser and version, it made sense to use a framework to hide the complexity and keep up to date. But with modern browsers this just isn't the case.
- take object of arbitrary depth {a: 1, b: [1, 2, {c: 3}]} and send as querystring parameters to Rails/Node/PHP/whatever backend
- take two parallel ajax calls and run some code when they're both done
- SPA w/ parameterized urls
- update the view when changes from a backend data source arrive, either via some sort of polling mechanism or a persistent connection.
- allow the user to edit fields on the selected item.
- automatically save any pending changes when navigating away from the page.
AKA every CRUD app ever. You are going to need a framework (most commonly MV*) to deal with this in a maintainable way. Whether you use a vendor or create your own, you will end up with a framework.
Virtual list controls that need to manage thousands of items/rows.
Hint: you're going to need a custom scroll bar.