That's all very nice, but it seems to do more with the fact that each person gets his/her own subdomain. And given that I found three xss's in about five minutes [edit: and then like ten more in the next five seconds, after realizing any input box works] doesn't give me confidence in their abilities.

http://www.pinkbike.com/news/search/?q=%3C%2Ftitle%3E%3Cscri...

http://www.pinkbike.com/product/compare/?items=466,%22%3E%3C...

http://www.pinkbike.com/photo/list/?date=all&text=%3C/ti...

http://www.pinkbike.com/buysell/list/?q=%3Cscript%3Ealert%28...

http://www.pinkbike.com/forum/search/?q=%3C/title%3E%3Cscrip...

Edit: I've stopped adding xss's. It's actually harder to find input boxes which don't lead to xss's than ones which do.

Whoever it was that fixed the XSS (necro?), I'm impressed how fast that was.

But please don't rely on just escaping < and >. You have to worry about double-quotes too, I can end a string and add a "onload" or "onfocus" attribute if it's already in a tag. And sometimes you have to worry about single quotes. In fact, there's a lot to take a look at.

Instead of just fixing the case at hand, try to be proactive about it. Check to make sure you don't have anything else.

Edit: Click the search box, for example. http://www.pinkbike.com/forum/search/?q=%22%20onclick=%22ale...

Yup. I actually wrote a filter to our framework when it was built years ago, but as new people help out on dev they don't always use the framework. My bad and I do appreciate the smack. I humbly bow to you with egg on my face.

Summary (he really should have said this part upfront):

He is (ab)using subdomains, by giving every user a different subdomain on his site. So a typical page can have links to hundreds of different subdomains on the page.

For a normal site disabling prefetching is not necessary or a good idea.

This article should be renamed: "Why using hundreds of unique subdomains is not a good idea." Besides all the DNS queries, you are also messing with caches.

I don't know anything about his site, but I don't see any reason that every user needs a unique subdomain.

One reason to do per-user domains is to prevent cookie leakage caused by user-specified content. A few years ago LiveJournal got bitten by a firefox bug that allowed JS execution inside CSS. Since they allowed each user to style their own page that meant that a malicious user could make the CSS at "www.livejournal.com/~evil_user" capture the viewer's www.livejournal.com cookies. That's why they moved everybody to username.livejournal.com, which previously had been a pay-only feature.

I think the really perplexing about this article is why browsers are so aggressive with DNS prefetching. If there are links on the page to 300 different domains, is there much benefit to precaching ALL of them? It seems clear that very few of the domains are likely to be clicked on. If anything they're probably hurting performance by bombarding the local nameserver with a flood of requests.

It sounds like browsers need a better precaching heuristic.

Unfortunately subdomains are not secure either, because subdomain is allowed to place cookie for the whole domain, performing session fixation attack:

http://skeptikal.org/2009/11/cross-subdomain-cookie-attacks....

the livejournal change:

http://www.davidpashley.com/blog/computing/livejournal-mozil...

> is there much benefit to precaching ALL of them? ... If anything they're probably hurting performance by bombarding the local nameserver with a flood of requests.

It's a tragedy of the commons. Each individual user gains personal performance from DNS prefetching. The gain for each individual user is greater than their loss to their share of the performance penalty imposed on the entire system.

The real super smart answer would be predictive prefetching based on pointer movement. When the browser sees the mouse heading towards a link, start prefetching. Or at least on mouseover.

That doesn't sound like a good enough reason, why not just use HTTPOnly cookies to prevent that?

Well, first off the 2006 LJ issue predates the implementation of HttpOnly cookies I think. They're a fairly recent invention.

HttpOnly doesn't really fix the issue though. It prevents evil js from straight-out stealing the cookie, but it can still make HTTP requests that will include the cookies. Therefore the script could still do bad things the user's behalf (add or remove friends, scrape private data, post spam, etc...)

The LJ story is just an example of a general lesson: if you can put user-generated content in a different domain (i.e. browser security context) than your authentication-critical cookies you are in a better position. That way any javascript leakage at evil-user.yourdomain.com can't harm your users any more than if they were viewing it at www.evil-users-domain.com.

It's a good point, and in retrospect I may not have used the username.example.com pattern, but this was a choice made may years ago. Users seem to like this, and unfortunately these recent changes in browsers make this pattern less desirable. Some other sites that do this subdomain style are the likes of tumblr.

Which is fine, as long as one doesn't display links to a lot of different users on one page (like an avatar alongside comments, hosted on the <user>.site.tld domain). As long as all those different user domains don't appear on one page, there's nothing to prefetch...

The real issue is that the cost per query is negligible here and DNS providers shouldn't charge by the query yet almost all of them do. They have to have some metric to separate the big guys from the little guys and charge the big guys more, but this is a crude way to do it.

Consider that publishers often have little control of how many DNS requests they get, so to charge for something out of your control seems utterly bizarre to me. Nice to see in this instance, publisher was able to make a meaningful improvement.

Also keep in mind, I used to run the largest free DNS service in the world so I'm well aware of what I'm talking about and am totally biased on these matters. :-)

The thing to note is that implementation of browser prefetching is further putting DNS requests out of the control of the publishers. 10 billion queries was something like 5000 dns req/s and that is considerable resources of some dns service. I wish dns was not per query costs, but it would seem that is the metric that resources would be based on. The point of the article is to let people know that if you fall into this dns pattern, it maybe a result of the recent prefetching done by browsers, and you may have a way to resolve the issue.

Yep -- nice to see that in this instance, as I noted. Then again, why should you be punished for your success if the reasons for query increase are based on simple traffic growth? Or worse, a competitor just sending you a ton of DNS traffic.

Is 5000 dns req/s really a considerable amount of resources? We were previously serving it on old hardware, with negligible (<5%) CPU usage (tinydns) and bandwidth.

If they have a large number of customers all doing 5000 req/s, perhaps they will need something a bit more powerful.

Perhaps a per-query charge isn't optimal, but how else can a provider of a service equate supply with demand?

More load implies additional cost, and DNS providers are no exception; there is an incremental cost of serving a DNS response. Granted, it's small, but real nonetheless.

One could charge based on bytes transferred instead, but charging for bytes is functionally equivalent to charging for responses.

I'm surprised the article doesn't actually tell you the meta http-equiv to disable DNS prefetching. It mentions that it helped out tremendously, though. Here it is:

<meta http-equiv="x-dns-prefetch-control" content="off" />

or, if you're more into HTML5:

<meta http-equiv="x-dns-prefetch-control" content="off">

I prefer HTTP headers instead of meta tags where possible.

An example solution at the load balancer level, assuming the use of Varnish:

In vcl_fetch:

  if (req.url ~ "\.(htm|html|php)" || req.url ~ "\/(\?.*)?$") {
    set beresp.http.disable-dns-prefetch = "1";
  }

In vcl_deliver:

  if (resp.http.disable-dns-prefetch) {
    remove resp.http.disable-dns-prefetch;
    set resp.http.X-DNS-Prefetch-Control = "off";
  }

do all the browsers that support the meta tag support consuming http headers for this "switch"?

Http-equiv is just a way of specifying an HTTP header in your HTML, so they really should.

Not closing the tag is nothing specific to HTML5, it just means you aren't strict on XHTML. HTML4 and others would be the same.

Well sure, but why does that matter? I could have said "or if you're more into HTML 4.01, 5, and anything before those versions of HTML..."

Pedantic.

I suppose it's just the irritation from the term HTML5 being used in all sorts of inaccurate or misleading ways. What you said wasn't inaccurate, but I feel like throwing in the term HTML5 was unnecessary. But yeah, it's pedantic. It's mostly me being irritated by others.

Before everyone goes rushing off to disable DNS prefetching, remember that DNS prefetching is generally a good thing that exists to make websites faster. And the faster your site is the more pageviews you can expect. Faster sites also have a lower bounce rate and better pagerank from Google.

Does Google take DNS prefetching into account when they measure the speed of a web site?

They get their timing data from the Google Toolbar, so yes.

This implies that providers are either severely limiting their caches, or expiring in a shorter than posted TTL. Even though pinkbike looks like it has thousands of users, one would expect the front pages to be largely identical for most user sessions, so the ISP dns caches should already have most of those username.domain.com records cached. Either that or the ISP's DNS servers are more numerous and distributed, with fewer customers each, or something like that.

Anybody at an ISP that can fill us in on DNS TTL mangling or cache limiting?

When I worked on Akamai's DNS server, I was told that most ISPs cached much more aggressively than we told them to. We would set TTLs at...I want to say 5 seconds...because we were directing traffic based on real-time network conditions; but most ISPs would cache for at least an hour.

Exact numbers may be off, of course; it was a few years ago.

See my reply here: http://news.ycombinator.com/item?id=2306788

He is creating thousands of unique subdomains, so nobody can cache anything.

username subdomains are a very popular feature with many websites, especially older ones. users like it too. there is nothing abusive about it. He cant scrap several years worth of links and throw everyone at www.example.com/username

And? Someone asked why the DNS wasn't caching and I explained.

Why can't this be handled with DNS wildcards? Surely they would allow caching on the nearest server.

I am actually rather shocked that people are being charged extra for DNS - surely the answer is to get any sort of cheap VPS and put DNS on that box? Then again, why are you depending for all aspects of your site's working at all, on a service which costs $2/month?

Even a 128MB RAM VPS could comfortably handle a huge number of requests.

Depends on the location of your users. Quicker to have someone host it is multiple places all over the world (assuming you care about perceived page load speed)

Create an A Record for your highest queried subdomain. Those will be cached and eventually decrease the number of queries. Depending on the number of subdomain you can create and remove it on signup and cancellation.

short answer for Firefox

   about:config 
   network.dns.disablePrefetch             (true)
   network.dns.disablePrefetchFromHTTPS    (true)

For those downvoting, did you want the server-side solution?

    <meta http-equiv="x-dns-prefetch-control" content="off" />

But that is a good thing for me!

ok i'm confused. why doesn't dns wildcard caching solve this?

DNS wildcards are server-side, so there's no way for the client to know that the response they got is for *.domain.com but not www.domain.com.

(I am not an expert on this, so may be wrong, but...) How does that change anything?

If the client needs to know foo.example.com they ask the nearest server. That will have the wildcard and will return the correct address. The nearest server is not going to be the root server for that domain, but something (typically) at the ISP. So the ISP DNS server will see more load, but not the root. Yet this article show the root as getting a huge number of hits. The only explanation I see is that they are not serving wildcard responses when they should be...

+ saved 10 billion DNS queries per month.

- wasted a total of 10sec of each user's time per month.

it's all about trade offs

In our (deviantART's) case, the 24-60 links to subdomains were set up so if you clicked them, they dynamically load with javascript on the client side, never actually hitting that subdomain. No time wasted on the user's side, since the browser was prefetching domains that were rarely being hit.

No, it's not; the issue is that the domains prefetch REGARDLESS of whether the links are ever clicked on. The browser is pre-fetching in case the user DOES click on the link. This really verges on being a browser bug, esp. since in firefox's case, it fetches both A and AAAA records.

The benefit is reaped in the event a link is actually clicked on. That's why browsers do it.

you have neglected to take into consideration the cost of running a hosted DNS. Costs for DNS service for any large traffic site is an important consideration.

Still continue to be about trade offs.... We could waste all day expanding my shallow example with items to weight in. but then we'd had to send a consulting invoice

As a fellow Mountain biker, I was surprised and glad that I saw pinkbike on HN.