Exciting! Meta supports shipping Zstd. We've been running an experiment on our side and have seen positive user metrics movement from switching to serving zstd to Chrome traffic. On Thursday, February 15, 2024 at 2:20:20 PM UTC-5 Mike Taylor wrote:
> LGTM3 > On 2/15/24 1:39 PM, Chris Harrelson wrote: > > LGTM2 > > On Wed, Feb 14, 2024 at 7:11 PM Nidhi Jaju <nidh...@chromium.org> wrote: > >> >> >> On Thu, Feb 15, 2024 at 12:50 AM David Benjamin <davi...@chromium.org> >> wrote: >> >>> On Wed, Feb 14, 2024 at 9:20 AM Yoav Weiss (@Shopify) < >>> yoav...@chromium.org> wrote: >>> >>>> LGTM1 >>>> >>>> On Wednesday, February 14, 2024 at 2:36:10 AM UTC+1 Nidhi Jaju wrote: >>>> >>>> On Wed, Feb 14, 2024 at 2:48 AM James Hartig <faste...@gmail.com> >>>> wrote: >>>> >>>> My employer ran into the window size during our pre-production >>>> validation and it was difficult to debug since it was working in cURL, the >>>> zstd CLI, and only presented itself on certain URLs. I appreciate Nidhi >>>> responding to our issue so quickly and updating Chrome to have a more >>>> helpful error message in the future. The Go package we use already updated >>>> their default <https://github.com/klauspost/compress/pull/913> to 8MB >>>> (without any awareness to Chrome's size) which should help future users of >>>> that package but there might be other packages out there that might not >>>> have a low enough default. The updated Chrome error message will help but >>>> only if you run into that error message when testing; which might not if >>>> you happen to be testing with small responses. I'm not sure where >>>> developers should be looking to be aware of the window size. Does it make >>>> sense to mention in the Chrome Status entry? If the spec is updated that >>>> might be good enough but I just wanted to discuss other avenues that might >>>> be more developer-aware. >>>> >>>> >>>> Thank you, I've included these details about the window size limits >>>> under the "Interoperability and Compatibility Risks" section in the >>>> ChromeStatus entry. >>>> >>>> On Tue, Feb 13, 2024 at 6:43 PM Yoav Weiss (@Shopify) < >>>> yoav...@chromium.org> wrote: >>>> >>>> >>>> >>>> On Tue, Feb 13, 2024 at 9:29 AM Nidhi Jaju <nidh...@chromium.org> >>>> wrote: >>>> >>>> >>>> >>>> On Tue, Feb 13, 2024 at 4:18 PM Yoav Weiss (@Shopify) < >>>> yoav...@chromium.org> wrote: >>>> >>>> Thanks for working on this!! This is extremely exciting! >>>> >>>> On Tue, Feb 13, 2024 at 1:11 AM Nidhi Jaju <nidh...@chromium.org> >>>> wrote: >>>> >>>> Contact emails >>>> >>>> nidh...@chromium.org >>>> >>>> Explainer >>>> >>>> https://docs.google.com/document/d/1aDyUw4mAzRdLyZyXpVgWvO- >>>> eLpc4ERz7I_7VDIPo9Hc/edit?usp=sharing >>>> >>>> Specification >>>> >>>> https://datatracker.ietf.org/doc/html/rfc8878 >>>> >>>> Design docs >>>> >>>> https://docs.google.com/document/d/14dbzMpsYPfkefAJos124uPrlkvW7j >>>> yPJhzjujSWws2k/edit?usp=sharing >>>> >>>> Summary >>>> >>>> Zstandard, or “zstd”, is a data compression mechanism described in >>>> RFC8878. It is a fast lossless compression algorithm, targeting real-time >>>> compression scenarios at zlib-level and better compression ratios. The >>>> "zstd" token was added as an IANA-registered Content-Encoding token as per >>>> https://datatracker.ietf.org/doc/html/rfc8878#name-content-encoding. >>>> >>>> Adding support for "zstd" as a Content-Encoding will help load pages >>>> faster and use less bandwidth, and spend less time and CPU/power on >>>> compression on our servers, resulting in reduced server costs. >>>> >>>> Blink component >>>> >>>> Internals>Network >>>> <https://bugs.chromium.org/p/chromium/issues/list?q=component:Internals%3ENetwork> >>>> >>>> TAG review >>>> >>>> https://github.com/w3ctag/design-reviews/issues/930 >>>> >>>> TAG review status >>>> >>>> Pending >>>> >>>> Risks >>>> Interoperability and Compatibility >>>> >>>> Servers that have a broken implementation of zstd might exist, but the >>>> risk of this is small. Additionally, middleware and middleboxes like virus >>>> checkers that intercept HTTPS connections might not support zstd, but >>>> might >>>> fail to remove it from the Accept-Encoding header in the request. >>>> >>>> Another known risk is interoperability between clients that support >>>> zstd regarding window frame sizes. In Chrome, we limit the window frame >>>> size to 8MB to prevent excessive memory usage, but this limit does not >>>> exist in curl and when using zstd directly. We have seen very few sites >>>> that use a window size larger than 8MB which causes decoding errors, but >>>> we >>>> have added new net error codes and debugging messages to help them >>>> understand what to do in this situation. >>>> >>>> >>>> I know we discussed >>>> <https://w3c.github.io/web-performance/meetings/2023/2023-09-TPAC/index.html#h.xn2d3li0b8op> >>>> >>>> this at length at the WebPerfWG. >>>> Can you summarize developments and/or findings since that discussion on >>>> that front? >>>> Should we expect the default output of CLI tools to be compatible with >>>> what we want to ship here? >>>> Should we expect interoperability between Chromium and e.g. curl? >>>> >>>> >>>> We've been discussing it with the zstd team at Meta at >>>> https://github.com/facebook/zstd/issues/2713. The plan is to take it >>>> to the HTTP WG at the IETF and either file an errata or publish a new >>>> document with more strict window size guidelines. The zstd CLI tool >>>> currently supports up to 8MB as a default, so the same limit. The library >>>> will use 128MB by default, however, and Curl currently supports up to >>>> 128MB >>>> windows. We expect those defaults to change to match any spec changes. In >>>> practice, we've seen very limited reports of sites running into this >>>> limit, >>>> and we've added helpful messages in Chromium to guide anyone who does run >>>> into it. >>>> >>>> >>>> Thanks! Pushing that limit into the standard and having curl (and other >>>> tools) follow that makes sense and seems important. >>>> >>>> Thinking out loud, the main risk here is for folks to be testing their >>>> content outside of Chromium (e.g. with curl) and then have that content >>>> break in Chromium. At the same time if content is tested in Chromium, it >>>> will work in another client that supports larger windows. >>>> So the (seemingly small) risk here is one we take on ourselves, rather >>>> than risk we externalize on the ecosystem. >>>> >>>> >>>> Yes, that sounds right. We'll continue to push to standardize this >>>> behavior across the ecosystem. >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> >>>> Gecko: Positive (https://github.com/mozilla/ >>>> standards-positions/issues/775) >>>> >>>> WebKit: Positive (https://github.com/WebKit/ >>>> standards-positions/issues/168) >>>> >>>> Web developers: Positive (https://crbug.com/1246971) Meta (Yann and >>>> Felix) and Akamai (Nic) are positive about zstd content-encoding on the >>>> browser. Meta has collaborated with us to improve the compression ratios >>>> for Meta origins during the experiment and is seeing positive user-level >>>> results. Alibaba is also supportive of shipping zstd support as they saw >>>> massive savings on their origins in terms of server CPU cost. >>>> >>>> Other signals: >>>> >>>> Ergonomics >>>> >>>> While both Zstandard and Brotli are clear wins over gzip >>>> content-encoding, which of Zstandard or Brotli to use depends on many >>>> factors, and site authors may need to experiment to identify the optimal >>>> choice for their content. >>>> >>>> Zstandard uses more memory for decompression than gzip. However, this >>>> is also true for Brotli, and we haven't seen any problems in practice. >>>> >>>> Activation >>>> >>>> The "zstd" Content-Encoding is not as widely supported by HTTP servers >>>> as gzip. Of the top 5 web servers, Nginx has a third-party module, which >>>> should also work for OpenResty (untested). Apache, IIS, and LiteSpeed >>>> appear to have no support. Explicit server support is often only necessary >>>> for dynamic content. For static (pre-compressed) content, Zstandard can >>>> often be supported just by configuration. >>>> >>>> Only one public CDN is known to be able to compress Zstandard itself, >>>> and some CDN's may require custom configuration to pass-through Zstandard >>>> correctly. >>>> >>>> Zstd support is not particularly difficult to implement for a server >>>> that already implements multiple content encodings. The C implementation >>>> has a straightforward API and there are implementations for many other >>>> languages. There is also a lively community of Zstandard enthusiasts which >>>> should help accelerate adoption. >>>> >>>> Security >>>> >>>> CRIME <https://en.wikipedia.org/wiki/CRIME> and BREACH >>>> <https://en.wikipedia.org/wiki/BREACH> mean that the resource being >>>> compressed can be considered readable by the document deploying them. That >>>> is bad if any of them contains information that the document cannot >>>> already >>>> obtain by other means. An attacker may provide correctly formed compressed >>>> frames with unreasonable memory requirements, and dictionaries may >>>> interact >>>> unexpectedly with a decoder, leading to possible memory or other >>>> resource-exhaustion attacks. It is possible to store arbitrary user >>>> metadata in skippable frames, so they can be used as a watermark to track >>>> the path of the compressed payload. It is important to note that these >>>> concerns apply to all compression formats, not just zstd. >>>> >>>> To mitigate these risks, similar to Brotli, we'll be advertising >>>> support for "zstd" encoding only if transferred data is opaque to proxies, >>>> to ensure that resources don't contain private data that the origin cannot >>>> read otherwise. >>>> >>>> >>>> I'm not sure what that means. Can you elaborate on that? >>>> >>>> >>>> This essentially means that, like Brotli, Zstd is only available in >>>> secure contexts >>>> <https://source.chromium.org/chromium/chromium/src/+/main:net/http/http_request_headers.cc;l=284-290;drc=4787fce6c51383f5631643ac3d14cc512d656de6> >>>> >>>> i.e. over https. >>>> >>>> >>>> Limiting zstd support to secure contexts makes perfect sense. However I >>>> believe the reason we're doing that for brotli is more around >>>> compatibility >>>> concerns with old network-based proxies >>>> <https://issues.chromium.org/issues/40930163> that aren't ready for >>>> non-gzip content-encodings. >>>> I don't think secure contexts do much to protect against BREACH if >>>> attackers can control parts of the response. At the same time, I don't >>>> know >>>> that we're doing anything on that front for other compression formats, so >>>> that seems fine. >>>> >>> >>> Right, secure contexts don't magically make dangerous features safe. The >>> *only* thing secure contexts do is make the name in the URL bar >>> meaningful. The user may still be talking to evil.example. >>> >>> It sounds like there are more risks discussed here than BREACH, so I >>> think we need to examine them separately: >>> >>> 1. Information leaks when you compress together attacker-controlled data >>> and secret data. (BREACH) >>> 2. DoS risks from the decompressor >>> 3. Watermarking from user-specific encodings of the resource >>> >>> For BREACH, the description of the document not being able to read it >>> confused me a little. When you compress something, the *length* of the >>> compressed resource, even when encrypted, gets leaked to all manner of >>> attackers via all manner of ways. I'm guessing the reference to the >>> document is that resource timing APIs allow the document to learn the >>> length of resources it otherwise cannot read? That is one attack vector >>> (not at all mitigated by secure contexts), but there are others. >>> Ultimately, BREACH means the server cannot *just* transparently >>> compress every resource it sends. In particular, any kind of dynamic HTML >>> resource will likely contain some attacker controlled strings. >>> >>> That said, the mitigation is mostly on the server to do. Once the >>> resource gets to us, the leak has already happened. The only connection to >>> proxies, and where we can do something on the client, is that sometimes >>> proxies will try to transparently compress all HTTP resources >>> indiscriminately. If that proxy is part of the network path and not the >>> site, it has no hope of mitigating this. So being opaque to proxies is good >>> and cuts out that minor component of the problem, but doesn't actually >>> address the broader issue. It's just fine because the broader issue is for >>> the server to address. (Though it means that our documentation to use it >>> should mention the server's responsibility here!) >>> >> >> Thank you for the additional discussion about the different security >> risks. I've added a note about the server's responsibility to the >> ChromeStatus entry to take care with including attacker-controlled data in >> compressed content. >> >> >>> >>> For DoS risks, secure contexts also don't do anything. We assume that >>> the attacker can direct the user to visit any website under their control, >>> so users could well visit https://evil.example and securely get a >>> DoS-triggering payload from it. As decompression happens in the network >>> service, shared across sites, that DoS would impact other sites too. So, in >>> order to deploy zstd or any such compression scheme, we need to mitigate >>> DoS attacks directly, usually by applying limits. It sounds like you all >>> have applied a frame size limit? Is that sufficient to avoid DoS, or are >>> the other avenues for a zstd decompression to consume excessive resources? >>> >> >> Yes, we added a window size limit of 8MB, which means that Chromium will >> use a maximum of 8MB memory buffer to decompress frames to protect it from >> unreasonable requirements. In addition, for zip bomb-like attacks, Chromium >> doesn't decompress faster than the renderer consumes data, so we won't >> accumulate excessive amounts of data in the network process. >> >> >>> Finally, the watermarking concerns also aren't mitigated by secure >>> contexts, but I think that's fine. This doesn't really apply to the web's >>> security model because we already assume the resource may be user-specific >>> in all manner of ways. (I mean, it can contain a Set-Cookie header!) >>> Rather, when we want two contexts to be uncorrelated, we control whether >>> they can communicate at all, rather than making assumptions on the encoding >>> mechanism. (Network state partitioning, cookie controls, etc.) >>> >> >> Agreed, for the content itself, we’ll continue to rely on the existing >> partitioning present in Chromium, as with other content encodings. >> >> >>> >>> Adding zstd to third_party/ in Chromium adds a large new code surface >>>> that processes untrusted data, which inevitably brings risks of new >>>> security holes. However, this is mitigated by the extensive fuzzing and >>>> security analysis done on zstd by Google and other community members. >>>> >>>> Furthermore, zstd is implemented in C, which is not a memory-safe >>>> language, and the network service is not yet sandboxed on all platforms. >>>> >>>> WebView application risks >>>> >>>> Does this intent deprecate or change behavior of existing APIs, such >>>> that it has potentially high risk for Android WebView-based applications? >>>> >>>> Apps which use a WebView to display content from Meta's servers will >>>> suddenly start using Zstandard. Since we've already extensively tested our >>>> implementation against Meta's servers in Chrome, no problems are expected. >>>> There is a killswitch. No special treatment should be needed. >>>> >>>> >>>> Debuggability >>>> >>>> No special support needed. >>>> >>>> Zstd content-encoding support is exposed to the devtools protocol, so >>>> developers are able to override it and view the headers from the inspector. >>>> >>>> A new net error has been added for decoding errors related to window >>>> frame size. >>>> >>>> Will this feature be supported on all six Blink platforms (Windows, >>>> Mac, Linux, ChromeOS, Android, and Android WebView)? >>>> >>>> Yes >>>> >>>> Is this feature fully tested by web-platform-tests >>>> <https://chromium.googlesource.com/chromium/src/+/main/docs/testing/web_platform_tests.md> >>>> ? >>>> >>>> Yes (https://wpt.fyi/results/fetch/content-encoding/zstd >>>> <https://wpt.fyi/results/fetch/content-encoding/zstd?label=experimental&label=master&aligned> >>>> ) >>>> >>>> Flag name on chrome://flags >>>> >>>> enable-zstd-content-encoding >>>> >>>> Finch feature name >>>> >>>> ZstdContentEncoding >>>> >>>> Requires code in //chrome? >>>> >>>> False >>>> >>>> Tracking bug >>>> >>>> https://bugs.chromium.org/p/chromium/issues/detail?id=1246971 >>>> >>>> Launch bug >>>> >>>> https://launch.corp.google.com/launch/4266275 >>>> >>>> Measurement >>>> >>>> https://chromestatus.com/metrics/feature/timeline/popularity/4629 >>>> >>>> Adoption plan >>>> >>>> In our experimental group, around 1% of responses use "zstd" >>>> content-encoding. Given the significant benefits of zstandard over gzip, >>>> we'd like to see it increase to 10% within 2 years. >>>> >>>> Estimated milestones >>>> >>>> Shipping on desktop >>>> >>>> 123 >>>> >>>> DevTrial on desktop >>>> >>>> 117 >>>> >>>> Shipping on Android >>>> >>>> 123 >>>> >>>> DevTrial on Android >>>> >>>> 117 >>>> >>>> Shipping on WebView >>>> >>>> 123 >>>> >>>> >>>> Anticipated spec changes >>>> >>>> Open questions about a feature may be a source of future web compat or >>>> interop issues. Please list open issues (e.g. links to known github issues >>>> in the project for the feature specification) whose resolution may >>>> introduce web compat/interop risk (e.g., changing to naming or structure >>>> of >>>> the API in a non-backward-compatible way). >>>> >>>> The current standard, RFC8878, doesn't require a limit on the window >>>> size used by HTTP servers when compressing Zstandard. An update of some >>>> form will be needed to ensure interoperability. >>>> >>>> Link to entry on the Chrome Platform Status >>>> >>>> https://chromestatus.com/feature/6186023867908096 >>>> >>>> Links to previous Intent discussions >>>> >>>> Intent to Prototype: https://groups.google.com/a/ >>>> chromium.org/g/blink-dev/c/GDsI0Hw-jYk/m/Yc5QZWD-AwAJ >>>> >>>> Intent to Experiment: https://groups.google.com/a/ >>>> chromium.org/g/blink-dev/c/I6IWfl95gRU >>>> >>>> >>>> This intent message was generated by Chrome Platform Status >>>> <https://chromestatus.com/>. >>>> >>>> >>>> -- >>>> You received this message because you are subscribed to the Google >>>> Groups "blink-dev" group. >>>> To unsubscribe from this group and stop receiving emails from it, send >>>> an email to blink-dev+unsubscr...@chromium.org. >>>> To view this discussion on the web visit https://groups.google.com/a/ >>>> chromium.org/d/msgid/blink-dev/CAMZNYAN7VRca4VfRqP7pi% >>>> 2BnqwDuor4ZVjF9yDNH1mZcXteQURw%40mail.gmail.com >>>> <https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAMZNYAN7VRca4VfRqP7pi%2BnqwDuor4ZVjF9yDNH1mZcXteQURw%40mail.gmail.com?utm_medium=email&utm_source=footer> >>>> . >>>> >>>> -- >>>> You received this message because you are subscribed to the Google >>>> Groups "blink-dev" group. >>>> To unsubscribe from this group and stop receiving emails from it, send >>>> an email to blink-dev+...@chromium.org. >>>> To view this discussion on the web visit >>>> https://groups.google.com/a/chromium.org/d/msgid/blink-dev/f823c2bc-f224-4ff7-9f78-e9eba9a4949cn%40chromium.org >>>> >>>> <https://groups.google.com/a/chromium.org/d/msgid/blink-dev/f823c2bc-f224-4ff7-9f78-e9eba9a4949cn%40chromium.org?utm_medium=email&utm_source=footer> >>>> . >>> >>> -- >> You received this message because you are subscribed to the Google Groups >> "blink-dev" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to blink-dev+...@chromium.org. >> To view this discussion on the web visit >> https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAMZNYAMDwUHi%3DLzPafbdg5CgQm4t27ZFLg3am53-no0D3Fg%2B%2BQ%40mail.gmail.com >> >> <https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAMZNYAMDwUHi%3DLzPafbdg5CgQm4t27ZFLg3am53-no0D3Fg%2B%2BQ%40mail.gmail.com?utm_medium=email&utm_source=footer> >> . >> > -- > You received this message because you are subscribed to the Google Groups > "blink-dev" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to blink-dev+...@chromium.org. > > To view this discussion on the web visit > https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAOMQ%2Bw-U7PCX0Yms-yTZo3uX2zRzjfV3q0nJ9%3DvG2oiYMQx4Nw%40mail.gmail.com > > <https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAOMQ%2Bw-U7PCX0Yms-yTZo3uX2zRzjfV3q0nJ9%3DvG2oiYMQx4Nw%40mail.gmail.com?utm_medium=email&utm_source=footer> > . > > -- You received this message because you are subscribed to the Google Groups "blink-dev" group. To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+unsubscr...@chromium.org. To view this discussion on the web visit https://groups.google.com/a/chromium.org/d/msgid/blink-dev/a33c485a-cd7f-4127-a97c-263c99b8c88bn%40chromium.org.
