News for hackers. Stuff that matters.
Re:375 Million pageviews for $15 a month by pdog on Wednesday October 03, 2012 @12:28PM (#4608230)
Seeing this (supporting over 10 million hits a day on an Amazon EC2 micro instance), it makes less and less sense to host a site entirely on Amazon S3.
Re:375 Million pageviews for $15 a month by spindritf on Wednesday October 03, 2012 @12:31PM (#4608250)
The advantage of hosting a site entirely on S3 is the minimal administrative overhead. You don't have to configure, run, maintain, update and secure a server.
Re:375 Million pageviews for $15 a month by awj on Wednesday October 03, 2012 @12:46PM (#4608336)
It depends heavily on your needs. Some top-of-my-head pros/cons:

EC2 Pros:

* Cheaper

* Easier to handle infrequently-updated content

* (Possibly) a simpler application pipeline

EC2 Cons:

* Have to administer a server

* Not durable, have to handle backups

* Worse availability (S3 goes down far less frequently than servers)

That said, if I'm reading it right 10M hits/day on S3 will cost you $100/day (edit: as aidos noted, I'm high by a factor of 10 here) in hits alone, so the price difference here obviously is not small.

Re:375 Million pageviews for $15 a month by aidos on Wednesday October 03, 2012 @12:56PM (#4608396)
I think you're a multiple out - it'll cost you $10 / day; still not to be sniffed at if you're trying to keep costs down. You pay the same for bandwidth with either solution.

We used s3 for a flash sale that even big boxes (with varnish) couldn't cope with. In our testing s3 didn't even blink (35,000 req/second). Real severs can't come close to that for the same price.

Re:375 Million pageviews for $15 a month by awj on Wednesday October 03, 2012 @01:09PM (#4608444)
Indeed I am, that's what I get for envelope math.

I think there's a price/performance sweet spot where a small server with something like varnish makes a lot of sense. Move outside of it in terms of hits/day or acceptable cost/hit and S3 gets a lot more attractive, especially when you can slap CloudFront in front of it so easily.

Actually, looking at the CloudFront pricing, unless you plan on doing a lot of cache invalidations it comes out much cheaper than S3.

Re:375 Million pageviews for $15 a month by aidos on Wednesday October 03, 2012 @02:34PM (#4608942)
Cloudfront in front of ec2 would be cheaper (you'd potentially pay a lot to have it in front of s3). Also, we discovered the hard way that cloudfront is nowhere near as resilient as s3 under load (and if you need to do a cache invalidation you are pretty much screwed).
Re:375 Million pageviews for $15 a month by awj on Thursday October 04, 2012 @11:46AM (#4612857)
Can you expand on the resiliency of cloudfront? We've had great success after switching to it for hosting of map tiles, but possibly haven't pushed it hard enough yet. (I pushed it hard, but defining "enough" is very, very tricky)

Cloudfront in front of S3 isn't particularly expensive. Invalidations aside they're billed the same way and you're paying less per bit|request.

Invalidations, however, are a bitch. We've had a couple of times that we put up a map, found a problem with it, then decided to wait out the CF cache rather than pay for a bajillion invalidations.

Re:375 Million pageviews for $15 a month by aidos on Monday October 08, 2012 @03:25AM (#4625870)
I have to be honest, I can't remember how much load we tried putting on cloudfront during our testing. We may have just assumed that as cf is a regular cdn and s3 is a big one of a kind data store, s3 was going to be more resilient. We did load test s3 and it scaled linearly (with some caveats).

For our use case (a single huge burst of requests) we were willing to drop all caching and pay the additional money. It was absolutely worth it (and under non-sustained load it didn't cost much).

We actually used s3 as a sort of cache in front of realtime data. So the client js was accessing objects on s3 that were being refreshed every second. It worked incredibly well where other things (like gae, who disabled our account under load) had failed us in the past.