TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

Does CloudFlare's Railgun really work? (Metrics)

10 pointsby dknechtalmost 12 years ago

2 comments

youngtaffalmost 12 years ago
Railgun&#x27;s an interesting idea but unfortunately the post doesn&#x27;t really give us enough information to really judge it&#x27;s effectiveness.<p>For example we don&#x27;t know enough about the &#x27;test&#x27; environment such as how much the pages vary from user to user, or from page to page which will have a large impact on deltas.<p>There&#x27;s also no detail on how other optimisations might perform e.g. serving all dynamic content through CDN so reusing CDN to origin connections, whether it might be possible to specify cache lifetimes for the dynamic pages etc.<p>There&#x27;s a short mention of some performance numbers at the start but then the article focuses on the reduction in outbound traffic (which is probably a good thing) so we&#x27;ve no idea whether Railgun, the other optimisation services that CloudFlare offer of just using a CDN is making these differences.
chrishollandalmost 12 years ago
I can answer some of these questions:<p>1) outbound traffic reductions are 100% driven by railgun. If you use cloudflare with railgun disabled, you&#x27;re using it much like a traditional dynamic acceleration product by Akamai or EdgeCast, and it does not reduce your outbound traffic, because the entire response payload is proxied out of your data center to the CDN&#x27;s edge. With railgun, the response body is crunched down to a diff. My post goes in good details into the figures we&#x27;ve seen with regards to off-peak organic long tail browsing and peak traffic driven by campaigns. Off-peak I&#x27;m down to about 1&#x2F;3rd to 1&#x2F;5th of my previous outbound consumption, and on-peak gets down to 1&#x2F;10th.<p>2) you can browse the site to see the various compositions of our page types: <a href="http://www.luxurylink.com/" rel="nofollow">http:&#x2F;&#x2F;www.luxurylink.com&#x2F;</a><p>Performance-wise:<p>3) I cannot apply any &quot;full-uri&quot; caching rules to any of my pages for various reasons: for one, any URI might send a completely different payload based on your device. You can bookmark a uri on your phone and open it up on your laptop after bookmark synchronization, and it always looks as beautiful as it should look for that device. Second, we&#x27;ve started to introduce dynamic merchandising tailored to every single user at every page load. Third, this is tied to our BI framework which does need every single request to fully go-through such that every page view continues to feed our learning algorithms.<p>2) with railgun disabled, performance-wise, I&#x27;ve measured an average 25% reduction in overall time to download a dynamic HTML document from new york and it speaks to the traditional WAN optimizations you&#x27;ve mentioned, which is on-par in features <i>and</i> performance gains with similar products I&#x27;ve used from Akamai and EdgeCast, for, you know, several grand a month base platform fee plus bandwidth metering. Except that cloudflare before railgun was only setting me back $20&#x2F;month plus $10&#x2F;month for 2 more sites (familygetaway and vacationist)<p>3) with railgun enabled the figure is now closer to a 35% reduction in download time from New York and 45% from London.<p>4) But as I alluded to at the end of the post, as far as what is felt by the end user, no railgun vs with railgun means perhaps a 100-200ms difference on the base HTML document total load time. And if you look at overall page load time, you&#x27;re going to spend several whole seconds downloading and rendering scripts, CSS, images. So perceived end-user benefit is likely limited. If they baked their compression into browsers I could imagine a performance improvement on-par with compression achieved, because the &quot;diff&quot; would travel all the way down to the client, instead of stopping at the CF proxy nearest to the client.<p>Still there are strong correlations between reduced HTML load time and increased Google crawl-rate on your site, which is a generally good thing.<p>Overall: Railgun has allowed me to insta-drop one whole bandwidth tier at my ISP, which is financially <i>material</i>. By the time our growth gets me back into the higher tier, i&#x27;ll be serving 5 times the audience I used to serve for that same amount of outbound consumption. We&#x27;re not cloud-hosted, but if we were, these bandwidth cost benefits would of course be a lot more linear.