</> Htmx – The Fetch()ening
htmx.org
htmx.org
> However, in order to keep my word that there will not be a htmx 3.0, the next release will instead be htmx 4.0.
technically correct.. the best kind of correct
Then 8
You can derive a lot of information about my age and current version from the dad-joke version string “I thought it was 6 7 when 7 8 9”
Source: I use both. :)
This is commendable, specially during times when libraries and programs aren't afraid of breaking changes and API churn.
I came from SPA-land and was tired of the fundamental architectural issue of having to keep the front-end and back-end state in sync.
I’ve compared Datastar and HTMX and decided on Datastar. There’s overlap between the two libraries in that they both support the request/response model, but with Datastar my learning investment takes me further and opens up new possibilities.
In one project I was able to remove a bunch of polling code and simply push a signal from the server to the browser when an external event occurred. The reduction in complexity was crazy.
On an internal tool I noticed I didn’t need Alpine.js anymore, and while anytime I can remove a dependency is a good time, the conceptual simplicity is what really makes me happy.
Now I’m doing a small app where I decided to make a streaming connection between browser and server and simply regenerate and send the entire view over that connection anytime something changes. Intuitively this felt wasteful but it turned out that with compression this works beautifully. There’s just less code in my app, and the code that’s there is less “fiddly” not having to deal with partial updates of the page.
If you’re coming from the world of SPAs, definitely check out both.
async def redirect_from_backend():
yield SSE.patch_elements('<div id="indicator">Redirecting in 3 seconds...</div>')
await asyncio.sleep(3)
yield SSE.execute_script('window.location = "/guide"')
[1] https://data-star.dev/how_tos/redirect_the_page_from_the_bac...
It’s not like synchronous servers e.g. Flask with Gunicorn (in the Python world) where you have a few pre-forked server processes. In that case it would be a waste to prevent one of those processes from serving other requests while sleeping, I agree.
Ultimately it’s cool though to be able to control the client from the backend asynchronously, updating signals (think of them as a kind of special front-end variable with superpowers) when something on the backend happens, or updating the client’s DOM, or running scripts.
This approach completely breaks sensible reasoning about what clients are responsible for. What's next, sending every keystroke to the server so it can tell the client to update a text input?!?
It still takes resources to keep SSEs open, load balance, reopen lost connections, etc. But worst of all is shifting fundamental logic back to the server for literally no benefit.
So confusing. I'm pretty sure it should be "inheritable", because "inherited" on an attribute means the attribute is inherited, not the element's children will inherit the attribute.
UPDATE: or "inherit", sounds like a command, little less confusing.
Like male pattern baldness, we are cursed by our parents to inherit it
That is one interpretation, but not the one I had reading it. I read it as "This target will be inherited by its children," rather than "This target is inherited from its parents," which, while a grammatical possibility, doesn't really make sense because you are specifying it right there with the equal sign.
Both "inherit" and "inheritable" have the same possible double readings. Something like "pass-to-children" or "pass-down" would remove the possible ambiguity, but I'm not sure like either better than "inherited".
Based on this section, it will be interesting to see how this evolves. I've used HTMX a bunch but after stumbling on Datastar I've come to prefer it. Partially because I don't need something like alpine.js to get some frontend goodies but also because I've come to appreciate the power of SSE streaming morphable targets to the browser
This feels like a repeat of the Python 3.0 strategy, though obviously at a much smaller scale. Some stuff is of course hard to roll out but to me it feels "obvious" that having a 4.0 with the inheritence change (or even better, a 2.1 with a toggle on that change!), then a 5.0 with some other changes, then a 6.0 with other changes... all feels way easier to manage than a 4.0 with all the changes at once.
We have version pinning! People who want 2.0 can have 2.0 "forever", so version numbers that go up very high are not actually a problem. Many releases can of course be a bit of a frustration from a release maker's perspective, but given that htmx is the way it is (ain't even getting type checking helping you out on any of this like you would with React!), having the gradual path seems way better.
"I think I've handled the 10 changes in between 2.0 and 4.0... but forgot the 11th change" is a constant annoyance with these huge major version bumps.
I will once again point to the Django strategy of having breaking changes be rolled out over several releases, including in between releases where "both" models exist. It is a very nice way to do release management when rolling things out, gives good checkpoints for pushing things out, and overall means users are seeing less bugs. Going from `XMLHttpRequest` to `fetch` really might not be a feasable thing to toggle, but a lot of the other stuff in that list feels like it.
Yeah this is a huge fallacy. If you do a breaking change, you want to do as few breaking changes at once as possible, with as little impact as possible and in the best case scenario you're in contact with every single person that is affected by the breaking change, so that it can be handled without anyone noticing anything other than having to bump a library version, which they were doing anyway since they're running a new Python version in the first place.
The dumbest change in python 3 was removing the print statement [0]. Maximum breakage, minimum gain. They could have introduced the new function as println() and deprecated the print statement in Python 3.
But, my thoughts immediately go to Datastar, which has Fetch, SSE, declarative signals and js expressions, dom morphing, and much more - in a tiny package. I find it to have a more flexible, expressive and standards-compliant API as well. And it'll soon have a simple reactive web components and css framework as well.
At this point, why use HTMX when it really seems like (a heavier) Datastar-lite?
But Datastar is different. The project is literally owned by a 501c3 non-profit. The devs have dayjobs and donate their spare time to this. Funds are for going to conferences or hosting their own
And 99% of the features/value that I mentioned is MIT licensed, and the "rugpulled" code is still available to easily port via the plugin API.
If the Postgres team released PGPro, swore it just contained anti patterns and you can just write the code yourself if you needed that feature, you’d roll your eyes, no?
It’s about the fact they went there, not about the intentions.
It doesn't put me off the project! I will happily pay for great software that enables me to make money just like I'd pay for good tools around the house (once I get started with D*).
I was just pointing out that in my mind the criticism is totally fair game because many OSS do not do this, that is all. By the looks of it you agree.
I think I want to just mention something here that is really important. The Datastar team have done more to push for performance, better UX and the ways of thinking around SPAs and MPAs than multi billion dollar companies have done in the last couple of years. Not forgetting HTMX, this cooperation has been incredibly fruitful and I am very excited about the future with you guys.
Moreover, the FOSS code still exists and would take 2 minutes to update to the current plugin API (I have Datastar pro and the code is almost exactly the same)
https://github.com/starfederation/datastar/blob/v1.0.0-beta....
I'm also sure this has already been explained in comments to other posts here as well.
I'm too tired to parse this logic, but I suspect it is a novel entry in techcorp doublespeak/dirty tricks.
* Datastar was re-written from the ground up, numerous times.
* They didn't want to update and maintain the plugins that they viewed as unnecessary/anti-patterns
* People wanted them still, so they said "fine, pay us to port it". Or, do it yourself - the MIT code is sitting right there and the changes are not all that significant. You'd also learn more about D* while at it. I linked in the parent comment to the MIT code - would not be difficult for anyone to do.
I suspect that in the long-run (probably not too far from now), they'll just make those plugins MIT again as the real value of Pro is the inspector, and soon their WIP web component framework (Rocket) and css framework (stellar) - all of which have always been being a commercial license.
p.s. there's no techcorp here. Its literally 3 guys with day jobs donating their time to a 501c3-registered non-profit. Funds go to things like going to conferences, or holding their own.
Ah, yes, a debugging tool. Only professionals need those.
Whatever the case, you dont truly need it, but it is helpful. You buy it for convenience as well as to support the project.
The Datastar authors are wrong about this. History push is a very important part of the hypermedia-driven application approach. Because URLs are super important. And we want to make sure that the correct URL is shown for the currently-loaded view, and that the view is reproducible given the URL (as much as possible) so that bookmarking and copy-pasting to send URLs just works as expected.
A really nice article came out about this just recently: https://alfy.blog/2025/10/31/your-url-is-your-state.html
I also wrote a bit more about it here: https://dev.to/yawaramin/why-hx-boost-is-actually-the-most-i...
As it turns out, I shared that very same article in the Datastar discord the other day! Here's some other good ones that I found while digging into the topic, for anyone who cares.
* https://warpspire.com/posts/url-design/
* https://blog.jim-nielsen.com/2023/examples-of-great-urls/
* https://www.w3.org/Provider/Style/URI
* https://www.hanselman.com/blog/urls-are-ui
I strongly agree that good urls are very important. But I don't see how D* prevents correct urls/history at all... You can click anchor links just fine for pages that are genuinely separate pages. If its just a sub page, filter etc, then i think in many cases it should only swap into the dom without updating the history.
Moreover, am I wrong to think that if you use hx-boost to swap in fragments, then the URL that gets updated/saved in history wouldn't actually load the same page if you loaded it from a bookmark? That wouldn't happen with non-boosted anchor links.
Anyway, I'm not the best person to take up this argument. If you are interested at all in some respectful debate on the topic, it would be great if you came by the datastar discord where there's definitely people who would be better able to engage with it. I'd be eager to observe from the sidelines
It's a common pattern with Django and template partials that you check if the request is an AJAX request, in which case you just load a partial template to swap into the existing DOM. Or if it's not an AJAX request, your server-side logic loads the full template.
A simple example would be a to-do list at http://example.com/todo/. Clicking on a task item would swap it into the DOM without a full page load, and then you'd update the URL and browser history to http://example.com/todo/my-task/. Then if you open that URL in a new session, your server side logic would render your page with the "my-task" already selected.
It depends. If it's a 'resource' (in the REST sense) then it should actually push the URL of the resource into history, because the URL should correspond to the currently viewed resource. This is exactly what I was talking about earlier, it's super important as a basic hypermedia principle.
> if you use hx-boost to swap in fragments, then the URL that gets updated/saved in history wouldn't actually load the same page if you loaded it from a bookmark
See amanzi's explanation or my blog post where I explain the same thing. With htmx we can easily check for the presence of a request header in the backend and serve the appropriate version of the resource: either a partial (fragment) rendering or a full page that contains the resource. I highly recommend reading my blog post: it's not a huge commitment and it will clarify these issues for you.
The reason to use htmx is that it has a simpler interface optimized for the majority use-case.
With htmx, you are largely tied to a request/reply paradigm. Something happens that triggers a request (e.g. user clicks a button, or some element scrolls into view), htmx sends the request, and then it processes the response. The htmx interface (`hx-get`, hx-trigger`) is optimized to make this paradigm extremely simple and concise to specify.
Datastar's focus (last I checked) is on decoupling these two things. Events may stream to the client at any time, regardless of whether or not they were triggered by a specific action on the client, and they get processed by Datastar and have some effect on the page. htmx has affordances for listening to events (SEE extension, new fetch support) and for placing items arbitrarily on the page (out-of-band swaps) but if your use-case is a video game or a dashboard or something else where the updates are frequently uncorrelated with user actions, Datastar makes a lot of sense. It's a bit like driving a manual transmission.
Delaney is fond of saying that there's no need for htmx when Datastar can technically do everything htmx can [0]. But I think this misses the point of what makes htmx so popular: most people's applications do fit within a largely request/reply paradigm, and using a library that assumes this paradigm is both simpler to implement and simpler to debug. As an htmx maintainer, I often encourage people to even use htmx less than they want to, because the request/reply paradigm is very powerful and the more you can adhere to browser's understanding of it, the more durable and maintainable your website will be [1].
[0] https://data-star.dev/essays/v1_and_beyond
[1] https://unplannedobsolescence.com/blog/less-htmx-is-more/
I think that even with req/resp morph leads to a simpler majority use case and that's what Turbo and Datastar have both shown. No?
v4 makes almost no changes to the interface, other than to flip inheritance to be off by default.
> I think that even with req/resp morph leads to a simpler majority use case and that's what Turbo and Datastar have both shown. No?
Although you can use the idiomorph extension for htmx, I personally don't think idiomorph is simpler, because there's an algorithm choosing what parts of the page get replaced based on the server response; I prefer to specify exactly what parts of the page get replaced in much simpler terms, a CSS selector, with `hx-target`.
Per [1] above, my style is minimize partial page responses wherever possible, so the ones that I do have are bespoke and replace a specific thing.
My interest in htmx is more on the coarse-grained aspects of its interface, not the finer ones, which is a consistent theme in my writings about it [0].
https://dev.37signals.com/a-happier-happy-path-in-turbo-with...
1. Datastar supports req/reply just fine - be it via normal text/html responses, or SSE (0, 1, or infinity responses) https://data-star.dev/reference/actions#response-handling. So, the crux of your argument is moot...
Moreover, if htmx's real value is ajax request/response, then why are you introducing SSE as a first-class citizen now?
2. Datastar has data-on, and various other attributes, that allow for triggering far more actions than just backend requests, from far more (any) events. I'm glad to see that htmx is now following suit with hx-on, even if it is apparently limited in capabilities.
3. Datastar can do OOB-swaps just fine - that's literally the core functionality, via (their own, faster) idiomorph.
4. Its a misnomer that Datastar is for video games etc - again, as described above, it can do all of the simple things that that HTMX can do, and more. And, again, why is HTMX introducing SSE if its so apparently unnecessary and unwieldy?
5. What makes htmx popular is that it was the first library to make declarative fragment swapping easy. And Carson is just a god-tier marketer. Its nice to see that he's now realized that Delaney was on to something when he wanted to introduce all of these v4 features to HTMX 3 years ago, but was (fortunately for us happy users) forced to go make Datastar instead.
6. We havent even talked about one of the key features - declarative signals. Signals are justifiably taking over all of the JS frameworks and there's even an active proposal to make them part of the web platform. D* makes them simpler to use than any of them, and in a tiny package.
I, Delaney, and all other D* users are grateful for HTMX opening this door. But I reiterate my original question - now that HTMX is becoming Datastar-lite, why not just use Datastar given that the powerful extras don't add any complexity and comes in a smaller package?
<button hx-get="/contact/1/edit">
And here's the datastar one, edited for parity: [1]
<button data-on:click="@get('/contact/1/edit')">
The htmx one is simpler. There's fewer mini-languages to learn and the API makes more assumptions about what you want. As you noted, Datastar has more generalized mechanisms that are certainly less clunky than htmx's if you lean heavily into more signals- or event-driven behavior, but for (what I believe to be) the majority use-case of a CRUD website, htmx's simpler interface is easier to implement and debug.(For example: you will see the response associated with the request in the browser network tab; I'm not sure if Datastar has a non-SSE mode to support that but it wouldn't be true for SSE.) To each their own.
As for "well then why implement X, Y, or Z," as the OP notes, refactoring to use fetch() means you get them largely for free, without compromising the nice interface. So why not?
Moreover, the fact that Datastar is more generalized is actually better - HTMX has vastly more (non-standards-compliant) attributes that you need to learn.
vs
https://data-star.dev/reference/attributes
> I'm not sure if Datastar has a non-SSE mode to support that but it wouldn't be true for SSE.) To each their own.
My first point was literally saying that it has non-SSE and linked to the docs. You're not even trying to be objective here...
> So why not?
Yes, I have no problem with these things being implemented in v4. In fact, celebrated it in my original post. I brought it all up because you were describing that all as needless complexity in Datastar, but now you're implementing it.
Also, most of Datastar can be trivially disabled/unbundled because its nearly all plugins. That is largely not the case for HTMX.
Thus far, you've simply strongly confirmed my initial hunch that HTMX v4 is unnecessary compared to Datastar.
Which is why HTMX is having to bolt on more gubbins. Because, although it's less characters to type its fundamentally complected and therefore less composable.
I'm sure you've already warched it but if you haven't I'd recommend Rich Hickey's talk Simple made Easy.
I was simply asking, though, what I might be missing now that HTMX is becoming a (heavier) Datastar-lite (no signals and more). Given that these changes were literally previously rejected by HTMX and caused Datastar to even come into existence, it seems wholly appropriate to be making the comparison here.
Also, I can't speak for others, but when I've brought up Datastar in other HTMX-centric discussions here and elsewhere, its only when someone asks about things like SSE, idiomorph etc... I always say that if you're looking for those features, you might prefer to just use D... Now that those features are native to HTMX, I suppose they can just stay with it. But you get even more for less weight by moving to D
The largely nonsensical and overly-defensive responses from HTMX's devs/supporters have only made it clear to me that D* is the appropriate choice here.
If you are making technical decisions about web programming based on HN threads though, best of luck to you.
I don't make any claims about the D* version of morph or idiomorph as I use neither of those.
<a id="contact1" href="/contact/1/edit" role="button">Edit</a>
Hhhmmm..., that's more verbose than both :-(
HTMX's 4's morph is almost a copy paste of Datastar's.
I'm very excited about htmx 4.0 and sounds like things are headed in a promising direction.
learned that trick from fixi.js
I thought I'd include an example of replacing fetch for anyone that come across this.
const originalFetch = window.fetch;
window.fetch = function(url, options) {
if (url.includes('/api/user')) {
const mockUser = {id: 1, name: 'John Doe', email: 'john@example.com'};
return Promise.resolve(new Response(JSON.stringify(mockUser)));
}
return originalFetch(url, options);
};
<button hx-get="/foo" hx-on:htmx:config:request="ctx.fetch = myCustomFetch">
Do It
</button>htmx 1 htmx 2 htmx 4
I look forward to version 8 in 2030.
(Meaning there always was a version 3! Just include both htmx1 and htmx2 onto your webpage!)
It seems odd to me to treat these as different properties on their own? might as well keep it as an extension if it is not going to be more deeply integrated and let htmx 4.0 be smaller? Possible I may not be understanding the use case though
We treat everything as integers server-side (because floating point is the devil) and defer all formatting to the client. Implicit cascade means I can change fmt-x="currency:USD:decimals:2" in one place and watch it ripple down the entire table. It's 'maddening' the way CSS is maddening - which is to say, it becomes second nature after you stop fighting it.
That said, I fully understand the support ticket burden. Maybe the real lesson is: implicit inheritance is great when you control the domain (like 'format all money the same way'), but terrible when people want to do arbitrary things at arbitrary levels. I will bear that in mind as I complete genx.software
"Couldn't find the requested release version 4.0.0-alpha."
It sounds like they are referring to hx-swap and not arbitrary javascript callbacks though, and in that case, I don't see why calling JSON.parse() inside htmx (when the content-type is json) is that big of a deal.
1. make an MPA 2. each page is a resource 3. keep a stream open on the current state of that resource 4. ship, touch grass, repeat
Uh, yes? They wrote a literal book about why they think this is important: https://hypermedia.systems/
you can replace the fetch() function used w/ an event callback, etc
should allow you to do pretty much anything w/o any hacks