CloudFront Kills the CDN Excuse

In August we wrote up the Olympics as a delivery masterclass and noted the pattern’s one barrier: content delivery networks came wrapped in enterprise contracts. Minimum commitments, sales cycles, pricing that started where mid-market budgets end.

Yesterday Amazon removed the wrapper. CloudFront, launched November 18, is a CDN sold like a utility. Pay per gigabyte delivered, no minimums, no sales call, switched on from the same self-service console as S3. A pilot takes about a day to stand up, and the results justify the headline: the excuse for mid-market sites going without edge delivery just died. Every asset a CDN serves is a request your server never sees.

What a CDN does, one more time with feeling

The principle, for readers joining the infrastructure series here. Your server sits somewhere, one somewhere, and every visitor everywhere pulls every asset across the full distance between them and it. Distance is latency. Latency is felt.

A CDN replicates your static assets to servers spread across many networks and regions. Images, stylesheets, scripts, video. Downloads too. Each visitor pulls from the nearest copy. Your origin serves each asset to the network once, and the network serves it to the world from next door.

Pages assemble faster everywhere, most dramatically for the visitors farthest from your rack. And your origin sheds the bandwidth-heavy grunt work to concentrate on the dynamic pages only it can build. It is the caching ladder’s geographic rung. Until now, it was the rung with a velvet rope.

The pilot, with numbers of the honest kind

What we did launch week, on a client site whose media had already moved to S3 per this spring’s recommendation. We enabled a CloudFront distribution in front of the S3 bucket, minutes of console work, repointed asset URLs, and measured.

Results, in the directional terms field tests deserve. Page-load improvements were visible everywhere and largest for West Coast and European test points, exactly as the geography predicts. The client’s origin is East Coast, and distance was the tax being refunded. Asset delivery times to the far test points improved by large fractions, not small ones. Origin bandwidth dropped in proportion to what moved.

Projected monthly cost at this client’s traffic: genuinely lunch money, at the published per-gigabyte rates. Total engineering: an afternoon, most of it verification rather than construction.

The candidate assets, in priority order for anyone replicating. Video and audio first, the heaviest and most latency-sensitive. Then image libraries: campus photography, product shots, the media galleries that dominate page weight. Then downloads like PDFs, annual reports, and forms. Then the stylesheets and scripts every page loads.

Work down the list until the origin is serving mostly what only the origin can serve. That is the finished state, and most sites reach it in an afternoon or two. Dynamic pages stay home. That is the origin’s job.

The honest limits

Where the launch product sits against the Akamai tier, because the comparison matters for expectations. CloudFront today is a straightforward static-asset CDN. No streaming-video protocols, though progressive download works fine and is what most institutional video needs. Live events at Olympic ambition remain enterprise territory.

Fewer edge locations than the incumbents: fourteen at launch against the giants’ thousands. That means ‘much closer to most visitors,’ not ‘adjacent to all.’ Beta-adjacent tooling: invalidating a cached asset before its expiry is currently awkward, so version your filenames (logo-v2.png) and sidestep the problem. And your assets need to live in S3, which for many teams is a feature, since that migration already made sense, but is a real prerequisite.

None of these limits touch the core use case. All of them explain why the enterprise contracts will keep existing for enterprise problems.

Setup notes for whoever gets the ticket

The mechanics, condensed for the implementer. A CloudFront ‘distribution’ maps to an S3 bucket as its origin. Creating one returns a delivery domain, and a friendlier address like assets.yourdomain.org is a CNAME record away.

Cache behavior is governed by expiry headers. Set generous lifetimes on assets, and adopt filename versioning as standard practice from day one. A changed asset is a new filename, deployed by reference, no waiting on caches to notice.

Test from somewhere that is not your office. Free web-based checkers from multiple regions do fine, so the improvement you report is the visitor’s, not the local network’s. And leave the arrangement documented: origin bucket, distribution settings, the versioning convention. The next person to touch this will otherwise re-derive it from mystery, and ‘infrastructure only one person understands’ is a risk category that consumes too much of everyone’s maintenance time.

Cost governance, the last mechanic. The pay-per-use meter runs both ways, so set up billing alerts and glance monthly. At institutional traffic the numbers stay small. The habit of watching them is what keeps ‘utility pricing’ from ever becoming ‘utility surprise,’ here or on any of the metered services piling up in your stack this year.

The pattern, third verse

Step back and this is the same graph we drew in April, with a new point plotted. Infrastructure keeps dropping down-market. Measurement went from five figures to free. Testing tools followed. Computing and storage became hourly rentals. Now edge delivery, the architecture that carried the Olympics, is available to a nonprofit at pennies per gigabyte.

Each drop follows the same sequence. Enterprise-exclusive capability, then a self-service utility version with honest limits, then a widening middle where ‘what the big sites do’ becomes what everyone should do. The practices we described in August as the pattern behind record-scale delivery are now, quite literally, a console checkbox.

The planning implication for 2009 budgets, which per last month’s triage are being written with unusual care: performance work keeps getting cheaper, and the ‘buy improvement while it’s discounted’ quadrant just gained an entry. An afternoon of asset offloading, at utility prices, for measurable speed on every page for every visitor. That is the kind of line item that survives any budget interview we can imagine. The velvet rope is down. Walk through it.

We will fold edge delivery into our standard launch checklist from here, alongside the caching ladder, the sitemap, and the analytics goals that previous years’ announcements made standard. That is how this industry actually progresses. Yesterday’s enterprise exotica becomes today’s checklist line, and the sites built by teams who keep the checklist current simply work better, year over year, for reasons their visitors never see. The checklist grew again this week. Ours is updated. Update yours.