Or just don't even bundle if you can serve your users with ES6 modules.

Performance with native ESM isn't great since you get cascading requests, unless you have very shallow dependency trees.

They work fine for everything that doesn't require immediate response on first load, and even the there are ways to address that - for instance special pages for entry points and/or loading indicators.

For general performance you can cache the modules using a service worker.

In any case they are fine for most use cases.

That's exactly what my team told me when we transitionned to an SPA. Turns out, splitting properly an SPA is complex and bundling is pretty annoying at the moment.

Perhaps I can help? Which bundler? What issues are you encountering?

We're using Vite.js/esbuild, the bundler takes 2.5GB of RAM locally.

I can offer suggestions on code-splitting, but bundler memory consumption gets into the weeds quickly.

Yeah we all collectively had a look at it and there's not a lot of solutions available. The bundler needs to bundle everything at some point and you are going to pay the SPA tax.

That's why I personally don't recommend SPAs, it introduce a whole other class of issues on top of what you usually deal with. Google makes it work with Gmail and Youtube but it's Google, they can throw infinite engineers at any problem and could make those in pretty much anything.

I agree there is a complexity cost for bundlers. Whether to pay it is an equation that roughly aligns with how robust and interactive your client offering needs to be. Applications probably, websites probably not.

We're a web application and even with those, I'm not sure the bundling is worth it. Customers in some regions complain about the high payload, we made sure the chunks are split properly but with continuous deployment, there's a lot of cache invalidation anyways.

Why? I think most modern bundlers and spa routers supports code splitting out of the box by just using dynamic imports, aka: `const myComponent = await import("/path/to/component")`

You still are going to build the whole thing locally anyways, even with vite.js, it takes 2.5GB of RAM on my machine.

I don't understand your concern with it and why this is a problem.

I haven't personally noticed the vite/esbuild memory consumption, but I also run build only once at the end before deployment or generating the prod artifacts.

It's a problem because it slows down my machine and each js file which is added makes the ram and cpu problem worse, it just doesn't scale.

We're using typescript so building only once a day isn't an option.

I'm sorry but that doesn't make sense to me (especially your last sentence regarding TypeScript).

During development vite has HMR and you don't need to run build after every change. This is no different when using SSR.

But I don't want to further argue. If you think that SPA is not suitable for your use case, that's ok.

Maybe I don't have the right wording, I'm not sure what it does exactly but running "vite dev" takes 2.5GB of RAM permanently.

It does reload changes very quickly so I guess it has some caching somewhere.

> But I don't want to further argue. If you think that SPA is not suitable for your use case, that's ok.

I personally never seen a use case which did not end up the way I'm describing, my 2 previous companies suffered the same issues.

I even warned my current company that there's a good chance that it would end up the same way and sure enough it did.

> It's a problem because it slows down my machine

I don't know what machine you're using, but filling 2.5GB of RAM was already taking less than 200ms back in 2018.

I don't suspect memory usage is an issue here.

It wouldn't be an issue if other work tools were taking absolutely nothing yes but as it is now, 16GB of RAM with a JS oriented stack isn't enough anymore.

And getting more RAM is just delaying the problem anyways since it takes more RAM every week.

In practice, the SPA is so slow and the user's device is so much slower than the datacenter that the "something" you see is a useless splash screen, and the actual content takes longer to arrive than just doing it SSR.

Reddit is probably the most familiar example of this, because it actually has multiple implementations (old.reddit, i.reddit etc) that let you see the difference!

I am wary of gleaning anything on the matter from Reddit. Their engineering decisions don't fit either the convention nor the needs of most web applications.

As for whether the initial paint is of value, that comes down to product and implementation. Nothing is faster than offline-first or local cache, which SSR does not have a good narraative for.

"at least seeing something" leads to jira or teams where you open a thing and then its entries slowly pop in and shift as they get loaded. An illusion of a site being there that can't be used because you can neither focus on the text nor use any of the inputs as they shift beneath your eyes.

That'd be a poor implementation. CSR can generate immediately interactive forms that don't shift. JIRA and MS are absolutely subpar irregardless of this matter.