Hacker News new | more | comments | ask | show | jobs | submit login
HTTP/1 should die; HTTP/2 can do everything HTTP/1 could, only faster (medium.com)
38 points by pgjones 29 days ago | hide | past | web | favorite | 52 comments

Actually, no. HTTP/2 cannot be easily read or written by a human with a telnet or s_client connection without some additional tools. It also can't be supported by a lot of old software without some additional layer of indirection. This may or may not be important to you, but it is a thing that HTTP/2 cannot do.

> HTTP/2 cannot be easily read or written by a human with a telnet or s_client connection without some additional tools

This is, of course, why we also don't use chunked encoding and compression with HTTP 1 rather than writing better tools. I mean, you even have the example in that sentence about how the solution for using TLS was not to say “it doesn't work with telnet” but to use s_client, ncat, socat, etc.

> It also can't be supported by a lot of old software without some additional layer of indirection

This part is true but it's also the classic legacy computing problem: old software which cannot be upgraded will increasingly have security and compatibility issues which favor putting it behind a proxy anyway. This should factor into the cost of choosing not to maintain those systems rather holding back the future.

How is HTTP/1.1 holding anything back?

HTTP 1 has a fixed feature set and a lot of warts. Nobody is working on HTTP 1.2, etc. so the only way anything is going to get better will be in the HTTP/2 series of development. At some point we need to ask whether skimping on maintenance is actually better than getting those improvements.

I'm not sure why you're being downvoted. The financial/resource burden of learning new technologies should be considered and weighed against the advantages it delivers over the old technology.

You're right! Technology should under no circumstances by adopted blindly. Costs and opportunities should always be weighed carefully.

Given that TLS is mention in-line as also breaking this point, I think it perhaps possible that some readers might be of the opinion that the decision-making process you call for has already been conducted. Since TLS breaks the standard of "it is human-eyeball-friendly with telnet/netcat", HTTP/2 is little different in the minds of many.

Your caution is right and apt. The history of computing is littered with technologies that were unready, unsuitable, or otherwise not fit for purpose when adopted.

Hopefully we all agree that maintainability is the most important quality of code (after successfully performing the feature that it was meant to implement.) Writing readable, modular code is key to getting the most out of developer time and preventing bugs. With that in mind HTTP/2 is a huge step backwards.

And for what? Slightly faster speeds? This is a non-problem. When I see a slow site and view the network performance to discover why (using tools that would most likely not work under HTTP/2) here is what I inevitably see:

* Massive video and JavaScript for ads

* JavaScript tracking code

* Massive HTTP headers with unnecessary cookies

* Unnecessary, bloated JavaScript libraries

* Unnecessary, bloated CSS files

* Gratuitous unoptimized images

* No attempt at optimizing HTTP caching settings

* No attempt at HTTP compression (gzip)

These will still be the main bottlenecks with HTTP/2.

HTTP/2 solves no problem and makes network application development more difficult by making the network calls opaque and gratuitously complex.

One important difference: HTTP/2 requires* encryption. This makes getting up and running for local development and small deployments more difficult.

* https://http2.github.io/faq/#does-http2-require-encryption "[...] currently no browser supports HTTP/2 unencrypted."

As such, we know this statement to be false: "With WebSocket HTTP/2 support there is nothing that HTTP/1.1 can do that HTTP/2 cannot"

I'm guessing they're not carving out a niche for self-signed certs? I think that would have been a fair balance, of setting a "this is a self-signed cert" instead of the current "YoUr WeBsIte IS UNSECURED" (which isn't actually true).

Accepting and indicating self signed would also be handy in Tor's onion sites, since the only companies that will issue certs require EV2 and all your info and a pile of money. And Onioncerts are pretty much only Facebook's area :/

> (which isn't actually true)

how would I discern your self-signed certificate from the one served by the person arp-spoofing the gateway in a Starbucks?

No cert chain for starters. The better answer here is some sort of "self signed standard data" for non-CA certs. But right now, testing in a NAT using HTTP2 is a ugly nonstarter.

> No cert chain for starters.

why would the person who arp-spoofs the gateway in the Starbucks not be able to fetch your self-signed cert chain and mint their own matching chain with the exact same metadata (but of course differing keys)?

Then what's your proposal for NATted, self-contained (no gateway), and Tor Onion networks?

Sure, if I have a public IP and DNS records pointing towards it, I'm served by LE or a multitude other vendors. But that's a small number of machines on any network.

If you have a public DNS entry then you can get a certificate from Let's Encrypt using the DNS verification method, even if the name points to a private IP address. DANE would also be an option if it (and DNSSEC) were more widely implemented.

Local name resolution (mDNS) could take a page from what Tor already does and encode the public key fingerprint into the domain name. If the key uniquely matches the domain (perhaps just the first part, e.g. b32-key-hash.friendly-name.local) then you automatically get the equivalent of domain validation. While, by itself, this doesn't prove that you're connected to the right domain, by bookmarking the page and visiting it only through that bookmark you would get the equivalent of trust-on-first-use. Browsers would just need to recognize this form of domain name and validate the key against the embedded hash instead of an external CA.

Dude, we don't use certs because we think they're cute. We use certs to guarantee cross-checking validation of resources. If you make me a certificate for your own website, I don't care how beautiful that cert is, it is also useless for cross-checking.

The only people who ever see the “this is insecure” page when it is actually a perfectly fine self-signed certificate know better.

If you special-case self signed certificates, at it looks any less threatening, you will eventually have that page presented to unsuspecting victims.

Yes and no, because it is complicated. This is a fairly naive example of HTTP/2 to illustrate a point, while most websites I am called in on to optimize the load time are not built like this. This example only shaved off around 60ms for 20 requests, if your website has 20 requests to load, you have ads or you have a resource problem.

HTTP/2 spec says it should not share a connection across a host/port combo, any content you have loaded on a CDN, or a your own cdn.mydomain.com will be a separate connection. The reason CDNs are faster is because they are closer(lower latency) or it is common and already cached in your browser.

HTTP/2 still suffers from latency and TCP Window sizes, so no your 8mb website will still be slow after you enable HTTP/2, you still have to push 8mb out to the client. If you have a site loading over 80 resources, concat and minify that first before asking your server admin to turn on HTTP/2.

HTTP/1 clients gets around some network latency issues by issuing more than one TCP socket, just like SFTP clients using more than one thread. Because it is hard to overload a single socket when your latency for your ACK packets is 200ms+. If this wasn't true, Google would not be spending the time on a UDP based version of HTTP. HTTP/2

Overall, lower you content size, lower the number of requests it takes to load your initial website, THEN turn on HTTP/2.

You have some IE6/HTTP 1 tuning advice which is no longer necessary in there.

> The reason CDNs are faster is because they are closer(lower latency) or it is common and already cached in your browser.

Shared CDN hit rates tend to be rather low and usually slower than self-hosting unless you're requesting a LOT of resources because the DNS + TLS setup overhead are greater than the eventual savings. Opening multiple connections was useful back when you had very low limits for the number of simultaneous requests but the browsers all raised that and HTTP/2 allowing interleaved requests really tilted that into net-loss territory.

The best way to do this is to host your site behind a CDN so that initial setup cost is amortized across almost every request and you can use things like pushes for the most critical page resources.

> If you have a site loading over 80 resources, concat and minify that first before asking your server admin to turn on HTTP/2.

Similarly, this is frequently a performance de-optimization because it means that none of your resources will be processed until the entire bundle finishes streaming and the client needs to refetch the entire thing any time a single byte changes. The right way to do this is to bundle only things which change together, prioritize the ones which block rendering, and break as much as possible out into asynchronous loads so you don't need to transfer most of it before the site is functional.

Most of my tests on CDNs show they are faster unless the server is hosted in Chicago for me. I have put anycast SSL termination in front of sites before that have also decreased the ssl connection handshake, which barely beats CloudFront or CloudFlare on a decent day. If I had the money I could use Akamai which I know my ISP has before it even hits the internet, no way I could make my web server faster than that.

It is not a performance de-optimization. If you have 8mb of JS, sure you should break that up, but you have 8mb of JS and that is your first true problem. I aim for about 400kb or less of JS, on a marketing website it may change once a quarter, so I really don't care and they are all first time users anyway. For a web application, I also don't care because if they are mobile they need that in one request, gziped to like 40kb, so the hours it would take to optimize for the single byte changed is still a performance budget of 80ms.

> Most of my tests on CDNs show they are faster unless the server is hosted in Chicago for me.

For you, with warmed caches. The problem is that most users aren't you and so when they follow a link to example.com their client makes that first DNS request for example.com, starts the connection and TLS handshake, etc. and then sees a request for e.g. cdnjs.com and does that same work again for a different hostname. If (in that example) you were hosting your site on CloudFlare you'd have the same work for the first connection but not the second because it'd already have an open connection by the time your HTML told the client it needed your CSS, JS, etc.

Here's an example: check out https://www.webpagetest.org/result/190125_F0_a1a180631ecebd8... and notice how often you see “DNS Lookup” and ”Initial Connection” times which are a significant chunk of the total request time — and ask yourself whether it would be better for the render-blocking resources to have those times be zero. Especially measure how that works on a cellular or flaky WiFi connection, which is closer to what most people experience in the real world.

As to the gigantic bundle of JavaScript, do some real browser testing and ask whether it's better to wait for the entire resource to be loaded before any code runs — and to repeat that entire transfer every time one byte changes — or only block on the portions needed to render the page. Yes, having 8MB of JavaScript is too much in aggregate but the solution is to use less and use it more intelligently, not slap it all into a big ball of mud.

Those DNS lookup times for Azure are in fact terrible, no DNS query should take 400ms. You can pre-lookup DNS using html header tags, but nothing fixes a slow DNS query like that, that is a failure of the Azure Edge DNS servers.

I do real browser testing, and have Real User Monitoring setup to prove the results. Page load times from 600ms to 1.2s reported from the clients browser. I measure and test our app on 3g connection including the high packet loss 4g signal I get at my house. The point of my comment was that HTTP2 is not a fix all issues with performance on a site. The Webpagetest result would not have loaded in 2 seconds if they switched to HTTP2. Its 400 resources is what is slowing it down, not http/1.

Look at the times for Google where they loaded jQuery: because it's using an external CDN, the initial connection time (80ms) is greater than the total transfer time (69ms) on a relatively fast connection. It would be faster to serve it from your own server and then you also don't have to deal with things like SRI to maintain the same level of security.

If bundling everything and rendering on the backend doesn't provide a better user experience than breaking everything and loading asynchronously, your website is already too fat and you should focus on fixing that first. The things that shouldn't be bundled are content images, and they don't block rendering.


I haven't seen more recent studies on this, but I do remember a lot of use cases and benchmarks I read where if your user base was on high-latency or unreliable internet connections (packet loss, even minor), HTTP/2 would be slower than 1. Here is a good summary of the issue:


Like you mentioned, HTTP is slowly moving towards UDP instead of TCP (QUIC protocol) to combat this.

Cloudflare covers it a bit here too: https://blog.cloudflare.com/the-road-to-quic/

I'm a fan of HTTP/2, and I think most people will benefit, but I really hate these kind of posts that only highlight the best case scenario to prove a point that has wide-ranging consequences. I can go on google and type "HTTP/2 is fast" and be reaffirmed in everything I thought about HTTP/2 -- I just did -- and almost every single blog mentioned zero downsides to using HTTP/2.

These are really great documents to reference, thank you for posting them. I'm going to send them to a few dev teams I work with to share the knowledge.

I rather believe HTTP/2 will die when HTTP/3 is available.

After all HTTP/1 is very simple to implement and already widely used and optimized. It is usable for most of cases. Plus, maybe in the future, CDN can serve HTTP/2 to client while use HTTP/1 to read the source.

And currently web browsers still need to send Upgrade request in HTTP/1 to know whether or not a unknown HTTP server supports HTTP/2. I guess this will still be true after HTTP/3 comes out (alt-svc).

Actually HTTP2 upgrade isn’t done with an “Upgrade:” request, but rather with TLS protocol negotiation.

Oh silly me. I've implemented my own HTTP/2 server according to RFC 7540[0]. I forgot in the real world web browsers just send "h2" TLS-ALPN.

[0] https://tools.ietf.org/html/rfc7540#section-3.2

Aha, this article is by the creator of Quart[0]! I'm a big fan - one of my favorite new python packages. It's essentially a super zippy, flask-compatible asyncio python server. I switched a flask app over recently and saw an immediate 10-20x throughput gain (the app is entirely io-bound). pgjones was a pleasure to work with when I had a few issues and had to contribute a few compat fixes as well. Thanks for the awesome package!

[0]: https://gitlab.com/pgjones/quart/

Wow, any chance you could write up the throughput gain? It would be great to see some real production numbers.

Thanks for the comments.

I’ve looked at quart and it looks interesting but is there much point to running an async webserver for APIs mainly reliant on db-access if we use SQLAlchemy for database connections?

I’m under the impression that since the db stuff still is sync/blocking we won’t win much by running an asgi server instead of WSGI.

Give peewee-async as shot if you'd like async DB access as well.


Looks like Gino is a new project trying to bake asyncio ORM on top of sqlalchemy core: https://github.com/fantix/gino

While interesting for greenfield development and experiments I'd say that it's most often a to big of a leap to change both ORM (with all db-access-code) and microframework at the same time for a well established project at 20KLOC+.

For sure, definitely not a good path for an existing app :) You can use the sqlalchemy-core via async via a few projects (https://github.com/RazerM/sqlalchemy_aio) but for the full ORM there's no option, as far as I know.

It's an internal dashboard, so that was a stress testing figure - afraid it wouldn't be a terribly interesting write up. It was still exciting for me, because I'd been looking for more or less flask-with-asyncio for quite some time.

This can't be true, can it?

I was under the impression that HTTP/2 used a persistent connection w/ multiplexing. This seems like it would be very nice in a web-browser to front-end situation, but what about for internal service calls? Seems like persistent connection between services would mess w/ common load balancing schemes.

There's nothing stopping you from having one connection per request with HTTP/2. You could build your software to simply have the same behavior as HTTP/1.1 with keep-alive

Time will show what unknown challenges and problems http/2 carries. So far, the protocol is studied mostly by google(and less by cloudflare), there is limited research by independent parties.

From the article, i see that the author heavily relies on the chrome devtools to demonstrate the performance benefit, relying on chrome connection statuses.

My spdy and http/2 tests in 2016 did not show much imrovment in perceived load speed for our e-commerce site. optimizing delivery(for us - caching the pre-rendered javascript components and pre-loading some ajax) yield better results. ymmv.

So, Websockets over HTTP/2 are only available in latest Firefox (or experimentally in Chrome, if you manually turn on a flag for it). Almost no servers support HTTP/2 Websockets, too.

Sorry, it is a little too early to switch.

If you want HTTP/2 to succeed, you're going to have to start making little wins. Find a tiny, easy market and get them to use it. Then find a giant customer (Google never counts) and get them to use it. If it seems more complicated, nobody's going to pick it up unless they have to.

The alternative is to create big sexy splash pages, create a lots of hype, and lie to people about how easy it is to implement. When they're finally caught up in the complexity of implementation, it'll be too late to back out.

Same can be said about IPv6 - and could have been said about IPv6 20 years ago. Still waiting...

That's a tricky comparison because using IPv6 required updates to the client, server, and every box in between whereas HTTP/2 only requires the endpoints to be updated and has a graceful degradation path in most cases. Unsurprisingly, it's already far more common than IPv6 because you don't have to go to every enterprise on the planet and tell them to fix things even their network team is afraid to touch.

In contrast, HTTP/2 rapidly hit much higher numbers thanks to Firefox and Chrome shipping support via automatic updates. When CloudFlare deprecated SPDY about a year ago they were already seeing adoption numbers just under 70%:


Speed. It is not everything.

Except websockets.

This is now possible, the article talks about what impact HTTP/2 WebSockets can make. (See also RFC 8441).


I think my articles in the past have been too dull; I think people would like to read something interesting and entertaining so I'm trying to improve in this regard. Was the article a good read or did the title put you off?

I am not the one you answered, but the title DID put me off, and therefore I didn't read the article (and I probably won't read it, even if it is good).

I loathe the tendency to pack emotion and opinion into everything, and exaggerate everything, just because people think they'll get readers that way. Please don't.

Folks generally click baity articles more, but are unhappy about the fact they were baited into clicking a low quality article. It's a dilemma - do you want people to be happy or do you want more readers?

Feedback on the article - haven't you chosen a pathological case of a single click triggering 20 API calls that don't depend on each other? Wouldn't it be simpler to batch these calls on the server?

It wasn't a criticism against you specifically, I meant that it's just pathetic that using titles like "X should die" is the only way to get people's attention these days, I know how hard it is to make people pay attention on the internet and it's just sad that provoking them is the only way to do it

but can h2 do websocket?

First sentence: "Recently WebSocket support has been added to HTTP/2 (in RFC 8441)"

Applications are open for YC Summer 2019

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact