Don’t use micro instances for real work
There are some use cases where micro instances work in production environments (such as MongoDb arbitrators etc) but running anything in production on AWS on anything less than a m1.large instance is setting you up for failure.
Micro and small instances can be used for a personal Web site or tiny blog, but nothing else. When I visually think of a micro or small instance, I envision a PC sitting under a desk performing operations that a server in a rack should be hosting.
Have servers on standby. Configure a large instance, stick it in an Elastic Load Balancer with your other web servers, and just stop it.... ()
I really don't consider one server stopped, on standby, as scaling AWS. In the cases of Web servers, you might need to provision 5-10 Web servers, and having a single stopped instance is not going to allow you to accomplish that.
The better approach would be to have a server image prepared that can auto-provision itself (multiple times when needed) into the role defined through user-data passed to the image.
Scaling Web servers is an n+1 operation, not a +1 operation.
Seriously, use RDS
Seriously, if you are doing any real MySQL work on AWS, DO NOT use RDS. The reality is that AWS doesn't nearly do the EBS tuning that you yourself can do (RAID-0, kernel tuning etc) on your own EBS based MySQL implementation. See:
Getting Good IO from Amazon's EBS:
http://orion.heroku.com/past/2009/7/29/io_performance_on_ebs...
Running MySQL on AWS:
http://aws.amazon.com/articles/1663
Finally, for MySQL slaves, use RAID-0 over the ephemeral disks which have dramatically better performance than EBS.
If you're on Amazon EC2 or similar, it becomes phenomenally cheap to throw hardware at the problem. Running a single medium instance is $0.175/hour. If you need need to go from 5 servers to 50 servers for two days, it's just $420 to get out of a 48 hour jam. Need to upgrade to the beefiest 68 GB, 28 ECU RDS instance while your'e at it? $125 for 48 hours.
Performance tuning for 100x load that may never come again is usually hard to justify, so just know what your outs are if you're lucky enough to be swamped with traffic.
However, just be sure that you know how to scale out quickly and you can launch new instances (hopefully automatically). On most days, our eCommerce business gets by on just a couple of servers, but we have our autoscaling rules in place that will spin up/down servers within a few minutes.
Be sure to test regularly. Having a server boot up and add itself to the load balancer with old code/configs can wreak havoc on your database.
I agree with allowing AWS to auto-provision N new instances when needed.
But from the article and with my financial hat on, when the large instance is stopped and left in a standby role, are you still paying for it?
It is important to remember that a stopped instanced is not guaranteed to be able to be started again. If you have a m1.large stopped, there is a chance that Amazon will not have any m1.larges available in that zone/region when you try to start back up your instance.
Also know that the private dns name will change when a stopped instance is started again, so anything configured to point at that particular instance will have to be reconfigured.
You aren't paying for the CPU time, only the S3/EBS storage space that the image occupies (which is negligible for any normal single instance/image)
There's one exception: if you've got "internal" servers that are only serving requests from you, or running occasional cron jobs (for monitoring and stuff like that), and you're sure that user traffic changes will never affect them, by all means go for a micro or a small. Those things aren't part of your publicly visible stack, so there's no sense wasting money.
Seriously, if you are doing any real MySQL work on AWS, DO NOT use RDS. The reality is that AWS doesn't nearly do the EBS tuning that you yourself can do (RAID-0, kernel tuning etc) on your own EBS based MySQL implementation.
Do you have any numbers on the performance difference between a well tuned EBS-based implementation and RDS?
Because speaking from experience, EBS based MySQL machines are always, without fail, the primary maintenance headache in any stack. They're always the machines that wake engineers up at 4 AM because of some EBS crappery or somesuch nonsense that necessitates intervention, and the promotion/backup/restore procedures can be touchy and error-prone. That's not even to mention the work involved in setting them up, testing, and getting them tuned right, which are serious stuff compared to the "pick a name, size, version, and launch!" process that RDS provides.
Most of what I could find seemed to suggest that the performance differences between RDS and well tuned EBS MySQL servers of comparable size were relatively small. For instance, from http://www.mysqlperformanceblog.com/2011/08/04/mysql-perform...: "My benchmarks generally showed that I was not able to outstrip RDS 5.1′s performance with the combination of stock MySQL 5.1, EC2, and a 10-volume EBS RAID."
If you think that's wrong and the difference is really much higher, I'd be interested to know.
Part of me is skeptical that any serious differences would persist for very long: if it's so easy to do this "right" so that you end up with a EBS+EC2 MySQL server that significantly outperforms RDS for general-purpose workloads, why wouldn't the RDS team set up RDS servers that way behind the scenes? I would think that Amazon should know how to efficiently utilize its own EBS volumes, right?
Ex 1: We had a request triggering an asynchronous job that would take 3-4 seconds of processor time. Several times, this job was being throttled, so even in a low traffic situation each time one of those requests came, a whole micro instance was consumed on it making it unavailable for anything else. Switching to small instances solved the problem.
Ex 2: When launching a new load balancer, due to some sort of wrong DNS/elastic-ip configuration, it's not uncommon to be hit hard by traffic coming for the last owner of the given ip. We've seen this happen several times in the past. Again, switching from 10+ micro instances to 2 small instances cleared the problem.
There's an interesting post by Greg Wilson addressing the throttling problem better than I could: http://gregsramblings.com/2011/02/07/amazon-ec2-micro-instan...
Also, we are running a fantasy soccer game with trading windows, so the server load near the trading window closure time (the last hour or so) is orders of magnitude higher than average load. Ideally we would scale up when the load demands it, but with the platform versioning issue unresolved we need to run our platform at "full scale" almost constantly (or scale manually which is a really bad idea when someone forgets to scale back up again for the weekend).
We're running in a WISC setup.
Overall this article has little to do with scaling on AWS. Running any sort of production website on a micro instance is pretty foolish, besides a blog or static website perhaps.
