（评论）

（评论）
(comments)

原始链接: https://news.ycombinator.com/item?id=41475124

当压缩文件进行网络传输时，与较大的数据包相比，没有必要发送多个较小的数据包。每组一个、两个或三个压缩数据包是一起确认的，而不是单独确认的。然而，使用 WebP 代替 gzip 进行图像压缩的好处不大，仅导致尺寸略有减小。例如，使用 gzip 压缩的 88KB 图像使用 WebP 压缩到 83KB，而 Brotli 则压缩到 69KB。这种大小差异可能看起来很小，但在传输时，两种情况都需要相同数量的数据包，因为它们超出了默认 TCP 数据包大小允许的最大数量。因此，尽管发送更少的数据包可能会节省时间，但由于加载、执行和获取 WebP 格式需要额外的 JavaScript 处理，因此并没有真正的改进。此外，由于浪费了处理能力，所声称的时间节省甚至可能是负的。此外，其他证明压缩方法性能存在显着差异的实例很少见。本地测试表明，XZ 的压缩效率比 gzip 高出大约 25%，Brotli 的性能与 XZ 类似。 Zstd 的性能也与 Brotli 相当，但其易用性和兼容性使其成为大多数实际应用的最佳选择。此外，Brotli 还存在构建系统缺失、文档缺乏和测试失败等问题，使其难以有效地融入主流发行版。最后，通过表单处理用户生成的内容时会出现潜在的安全问题，因为特定的字节序列可能会操纵输出的 HTML，从而可能注入有害的脚本。

> the longest post on my site, takes 92 KiB instead of 37 KiB. This amounts to an unnecessary 2.5x increase in load time

Sure, if you ignore latency. In reality it's an unnecessary 0.001% increase in load time because that size increase isn't enough to matter vs the round trip time. And the time you save transmitting 55 fewer KiB is probably less than the time lost to decompression. :p

While fun, I would expect this specific scenario to actually be worse for the user experience not better. Speed will be a complete wash and compatibility will be worse.

That's certainly reasonable if you optimize only for loading time (and make certain assumptions about everybody's available data rate), but sometimes I really wish website (and more commonly app) authors wouldn't make that speed/data tradeoff so freely on my behalf, for me to find out after they've already pushed that extra data over my metered connection.

The tragedy here is that while some people, such as the author of TFA, go to great lengths to get from about 100 to 50 kB, others don't think twice to send me literally tens of megabytes of images, when I just want to know when a restaurant is open – on roaming data.

Resource awareness exists, but it's unfortunately very unevenly distributed.

There is an interesting "Save-Data" header to let a site know which makes sense to optimize for on connection but it seems to be Chrome only so far https://caniuse.com/?search=save-data

I wish there was a bit of an opposite option - a "don't lazy/partially load anything" for those of us on fiber watching images pop up as we scroll past them in the page that's been open for a minute.

and you'd need AI to tell you whats in the pictures because lots of restaurant sites just have photos of their menu and some designer with no web knowledge put their phone number, address, and hours in an image designed in Photoshop

It’s not just decompression time. They need to download the whole thing before decompression, whereas the browser can decompress and render HTML as it’s streamed from the server. If the connection is interrupted you lose everything, instead of being able to read the part you’ve downloaded.

So, for any reasonable connection the difference doesn’t matter; for actually gruesomely slow/unreliable connections where 50KB matters this is markedly worse. While a fun experiment, please don’t do it on your site.

Other major issues that I had to contend with:

1: browsers choose when to download files and run JavaScript. It is not as easy as one might think to force JavaScript to run immediately as high priority (which it needs to be when it is on critical path to painting).

2: you lose certain browser optimisations where normally many things are done in parallel. Instead you are introducing delays into critical path and those delays might not be worth the "gain".

3: Browsers do great things to start requesting files in parallel as files are detected with HTML/CSS. Removing that feature can be a poor tradeoff.

There are a few other unobvious downsides. I would never deploy anything like that to a production site without serious engineering effort to measure the costs and benefits.

It adds unicode characters before elements with the given class. Then it's up to the font to display those Unicode characters — in this case, based on the class names, one can infer that the font assigns an icon to each character used.

That makes sense, thank you!

So the purpose is effectively to have human-readable CSS class names to refer to given glyphs in the font, rather than having stray private use Unicode characters in the HTML?

Yep

This is a reasonable approach if you have a large number of icons across large parts of the site, but you should always compile the CSS/icon set down to only those used.

If only a few icons, and the icons are small, then inlining the SVG is a better option. But if you have too many SVGs directly embedded on the site, the page size itself will suffer.

As always with website optimization, whether something is a good option always “depends”.

More reasonable than this class+CSS would be e.g. a React/static-website-template/etc custom element that outputs the correct glyph. The output doesn't need to contain this indirection, and all of the possibilities.

seriously, why can't modern browsers turn off features like remote fonts, webrtc, etc. in settings. I hate when reading a bit then the font changes. Not to say fingerprinting risks.

Its a convenience packaging issue. An icon font is simply more convenient to handle. tags for a hundred images requires more work.

Icon fonts are used all over the place - look at the terminal nowadays. Most TUI's require an icon font to be installed.

Definitely fair and I was a bit harsh. It just seemed a bit nonsensical to go through such efforts to get rid of a few kilobytes while serving a massive font file. But I also understand that it's mostly for fun :)

That size difference is large enough to make a difference in the number of round trips required (should be roughly one fewer roundtrip with any sensible modern value for the initial congestion window).

Won't be a 2.5x difference, but also not 0.001%.

You don't need a new roundtrip for every packet. That would be devastating for throughput. One vs two vs three file packets get acked as a batch either way, not serially.

Also when you get to the end, you then see

> The actual savings here are moderate: the original is 88 KiB with gzip, and the WebP one is 83 KiB with gzip. In contrast, Brotli would provide 69 KiB.

At 69 KiB you're still over the default TCP packet max, which means both cases transmit the same number of packets, one just has a bunch of extra overhead added for the extra JavaScript fetch, load, and execute.

The time saved here is going to be negligible at best anyway, but there looks to be actually negative because we're burning time without reducing the number of needed packets at all.

Those numbers are for a different page. For the original page, the article quotes 44 kB with this method vs. 92 kB for gzip.

> At 69 KiB you're still over the default TCP packet max, which means both cases transmit the same number of packets,

What? No, they absolutely don't transmit the same number of packets. Did you mean some other word?

I expect what GP meant is the default TCP window size, so in a situation where bandwidth costs are dwarfed by roundtrip costs, these two cases will end up taking essentially the same time, because they will incur the same number of ACK roundtrips. Don’t know if the numbers work out, but they at least sound plausible.

No, there is no way the numbers would work out to the same number of roundtrips. The sizes are different by a factor of 2.5x, and the congestion window will only double in a single roundtrip. The only way the number of roundtrips would be the same is if both transfers fit in the initial congestion window.

They were probably thinking of the max size for packets in TCP, which is 64K (65535 bytes).

However, Ethernet has a MTU (Maximum Transmission Unit) of 1500 bytes. Unless jumbo frames are used.

And so I agree with you, the number of packets that will be sent for 69 KiB vs 92 KiB will likely be different.

Interesting, how does that add a round trip? For the record here's what I believe to be the common definition of an additional "round trip", in a web development context:

  - client requests X
  - client gets X, which contains a reference to Y
  - therefore client requests Y

So you're starting a new request that depends on the client having received the first one. (although upon closer inspection I think the technique described in the blog post manages to fit everything into the first response, so I'm not sure how relevant this is)

Unless a resource is very small, it won't be transmitted in a single atomic unit. The sender will only send a part of it, wait the client to acknowledge having received them, and only then send more. That requires a network roundtrip. The larger the resource, the more network roundtrips will be required.

If you want to learn more, pretty much any resource on TCP should explain this stuff. Here's something I wrote years ago, the background section should be pretty applicable: https://www.snellman.net/blog/archive/2017-08-19-slow-ps4-do...

In reality it's more like:

  - client requests X
  - server sends bytes 0-2k of X
  - client acknowledges bytes 0-2k of X
  - server sends bytes 2k-6k of X
  - client acknowledges bytes 2k-6k of X
  - server sends bytes 6k-14k of X
  - client acknowledges bytes 6k-14k of X
  - server sends bytes 14k-30k of X
  - client acknowledges bytes 14k-30k of X
  - server sends bytes 30k-62k of X
  - client acknowledges bytes 30k-62k of X
  - server sends bytes 62k-83k of X
  - client acknowledges bytes 62k-83k of X
  - client has received X, which contains a reference to Y
  - therefore client requests Y

It's all about TCP congestion control here. There are dozens of algorithms used to handle it, but in pretty much all cases you want to have some kind of slow buildup in order to avoid completely swamping a slower connection and having all but the first few of your packets getting dropped.

Not just modern. This was even more significant on slow connections, so they've kind of always done that. One could even argue that HTML, HTTP (specifically, chunked encoding) and gzip are all intentionally designed to enable this.

Seriously, if you're saving less than a TCP receive window's worth of space it's not going to make any difference to latency.

I suppose it could make a difference on lossy networks, but I'm not sure.

If the blob contains requests (images, but also stylesheets, JS, or worst case fonts), it will actually instead be a net negative to latency. The browser's preload scanner begins fetching resources even before the HTML is finished being parsed. That can't happen if the HTML doesn't exist until after JS decodes it. In other words, the entire body has become a blocking resource.

These are similar conversations people have around hydration, by the by.

Why is there more latency?

Edit: Ah, I see OP's code requests the webp separately. You can avoid the extra request if you write a self-extracting html/webp polyglot file, as is typically done in the demoscene.

It takes more time for your message to get back and forth between your computer and the server than it takes for the server to pump out some extra bits.

Even if you transmit the js stuff inline, the op's notion of time still just ignores the fact that it takes the caller time to even ask the server for the data in the first place, and at such small sizes that time swallows the time to transmit from the user's perspective.

That's fine, but if you're evaluating the amount of time it takes to load a webpage, you cannot ignore the time it takes for the client request to reach your server in the first place or for the client to then unpack the data. The time saved transmitting such a small number of bits will be a fraction of the time spent making that initial request anyway. That's all I'm saying.

OP is only looking at transmit size differences, which is both not the same as transmit time differences and also not what the user actually experiences when requesting the page.

Actually, I was rather wondering about that claim, because it seems accidentally cherry-picked. Regarding that post:

> This code minifies to about 550 bytes. Together with the WebP itself, this amounts to 44 KiB. In comparison, gzip was 92 KiB, and Brotli would be 37 KiB.

But regarding the current one:

> The actual savings here are moderate: the original is 88 KiB with gzip, and the WebP one is 83 KiB with gzip. In contrast, Brotli would provide 69 KiB. Better than nothing, though.

Most of the other examples don't show dramatic (like more than factor-of-2) differences between the compression methods either. In my own local testing (on Python wheel data, which should be mostly Python source code, thus text that's full of common identifiers and keywords) I find that XZ typically outperforms gzip by about 25%, while Brotli doesn't do any better than XZ.

XZ was never considered to become a compression algorithm built into web browsers to start with. Brotli decoder is already there for HTTP, so it has been proposed to include the full Brotli decoder and decoder API as it shouldn't take too much effort to add an encoder and expose them.

Also, XZ (or LZMA/LZMA2 in general) produces a smaller compressed data than Brotli with lots of free time, but is much slower than Brotli when targetting the same compression ratio. This is because LZMA/LZMA2 uses an adaptive range coder and multiple code distribution contexts, both highly contribute to the slowness when higher compression ratios are requested. Brotli only has the latter and its coding is just a bitwise Huffman coder.

> Why readPixels is not subject to anti-fingerprinting is beyond me. It does not sprinkle hardly visible typos all over the page, so that works for me.

> keep the styling and the top of the page (about 8 KiB uncompressed) in the gzipped HTML and only compress the content below the viewport with WebP

Ah, that explains why the article suddenly cut off after a random sentence, with an empty page that follows. I'm using LibreWolf which disables WebGL, and I use Chromium for random web games that need WebGL. The article worked just fine with WebGL enabled, neat technique to be honest.

It isn't neat as long as it doesn't work with all modern web browsers (even with fingerprinting protection) and doesn't have a fallback for older browsers. WWW should be universally accessible and progressively enhanced, starting with plain HTML.

I don’t support LLM companies stealing content and profiting from it without contributing back. But if you’re going to fight that by making things more difficult for humans, especially those with accessibility needs, then what even is the point of publishing anything?

I work on Batch Compress (https://batchcompress.com/en) and recently added WebP support, then made it the default soon after.

As far as I know, it was already making the smallest JPEGs out of any of the web compression tools, but WebP was coming out only ~50% of the size of the JPEGs. It was an easy decision to make WebP the default not too long after adding support for it.

Quite a lot of people use the site, so I was anticipating some complaints after making WebP the default, but it's been about a month and so far there has been only one complaint/enquiry about WebP. It seems that almost all tools & browsers now support WebP. I've only encountered one website recently where uploading a WebP image wasn't handled correctly and blocked the next step. Almost everything supports it well these days.

The key note for understanding the approach (without diving into how to understand it's actually wrangled):

> The only possibility is to use the WOFF2 font file format which Brotli was originally designed for, but you need to make a whole font file to leverage this. This got more complicated recently by the fact that modern browsers sanitize font files, typically by the OpenType Sanitizer (OTS), as it is very insecure to put untrusted font files directly to the system. Therefore we need to make an WOFF2 file that is sane enough to be accepted by OTS _and_ has a desired byte sequence inside which can be somehow extracted. After lots of failed experiments, I settled on the glyph widths ("advance") which get encoded in a sequence of two-byte signed integers with almost no other restrictions.

Fantastic idea!

Chromies got in the way of it for a very long time, but zstd is now coming to the web too, as it’s finally landed in chrome - now we’ve gotta get safari onboard

Brotli and zstd are close, each trading in various configurations and cases - but zstd is muuuuuch more usable in a practical sense, because the code base can be easily compiled and linked against in a wider number of places with less effort, and the cli tools turn up all over for similar reasons. Brotli like many Google libraries in the last decade are infected with googlisms. Brotli is less bad, in a way, for example the C comes with zero build system at all, which is marginally better than a bazel disaster of generated stuff, but it’s still not in a state the average distro contributor gives a care to go turn into a library, prepare pkgconf rules for and whatnot - plus no versioning and so on. Oh and the tests are currently failing.

Realistically, everything should just support everything. There’s no reason why every (full featured) web server and every (full featured) web browser couldn’t support all compelling data compression algorithms.

Unfortunately we live in a world where Google decides to rip JPEG-XL support out of Chrome for seemingly no reason other than spite. If the reason was a lack of maturity in the underlying library, fine, but that wasn’t the reason they offered.

> There’s no reason

Of course, there is - and it's really boring. Prioritisation, and maintenance.

It's a big pain to add, say, 100 compressions formats and support them indefinitely, especially with little differentiation between them. Once we agree on what the upper bound of useless formats is, we can start to negotiate what the lower limit is.

It’s not prioritization of the code - it’s relatively little implementation and low maintenance but security is critical here and everyone is very afraid of compressors because gzip, jpeg and various others were pretty bad. Zstd, unlike lz4 before it (at least early on) has a good amount of tests and fuzzing. The implementation could probably be safer (with fairly substantial effort) but having test and fuzzing coverage is a huge step forward over prior industry norms

I qualified it with compelling which means only including formats/encodings which have demonstrably superior performance in some non-trivial respect.

And I qualified it with mature implementation because I agree that if there is no implementation which has a clear specification, is well written, actively maintained, and free of jank, then it ought not qualify.

Relative to the current status quo, I would only imagine the number of data compression, image compression, and media compression options to increase by a handful. Single digits. But the sooner we add them, the sooner they can become sufficiently widely deployed as to be useful.

While peeking at the source, I noticed that the doctype declaration is missing a space. It currently reads , but it should be

That page breaks my mouse gestures add-on! (Or, I guess we don't have add-ons anymore but rather something like script injections that we call extensions, yay...) Interesting approach to first deliver 'garbage' and then append a bit of JS to transform it back into a page. The inner security nerd in me wonders if this might open up attacks if you would have some kind of user-supplied data, such as a comment form. One could probably find a sequence of bytes to comment that will, after compression, turn into a script tag, positioned (running) before yours would?

Yeah that's plausible, you definitely don't want any kind of untrusted data in the input.

Something I wanted to do but clearly never got around to, was figuring out how to put an open-comment sequence (

（评论） (comments)

（评论）
(comments)