More information about the Underscore mailing list

[_] AWS users

Tim Wintle timwintle at gmail.com
Tue Feb 5 15:50:43 GMT 2013

On Tue, 2013-02-05 at 14:47 +0000, Andrew Holway wrote:
> Could you describe your setups? Joys and sorrows?

My use is probably quite unusual - we own/run the vast majority of our
own machines and services, but use a touch of AWS.

S3:

We use S3 for some small files, essentially as a shared filesystem
service that multiple frontends can write to.

We've had several issues with it not meeting our (very high) uptime
requirements - we're still talking 99.9% uptime, and we're hoping to
move away from it to some kind of filesystem/SAN on our own network.


EC2:

We occasionally use EC2 to start front-end servers (essentially just
caching proxies) in locations where we have sudden requirements for
performance, or where we've had networking issues.

For example, a few hours before Sandy hit NYC we shifted our US traffic
from our servers in New Jersey to EC2 instances on the west coast.

We did something similar last year when there was an ongoing routing
issue on a major link between networks in central/south America and the
US, and we started some instances in Brazil so we could route traffic
back along known good locations.

We don't get much notice about these things and they're normally
short-lived, so EC2 is very cheap for this use case.

(We've optimised the whole process fairly well - it takes around 10
minutes to start an fresh instance, install our entire stack, reboot the
machine and register services for monitoring/server alerts etc, if it's
on EC2 or raw iron)

If you know your load requirements in advance then I'd say EC2 is
generally an expensive option though (with the exception of Sydney where
I'm still amazed with the price they get it to!). I believe they've just
dropped their prices, so it's worth looking again.


If you're doing SOA stuff then last time I tried it the networking
between instances was a killer for low-latency systems - certainly don't
consider yourself on the same rack as your other instances, and expect
wide fluctuations. This may have got better but I don't think so.


Glacier:

We're tentatively looking at using Glacier for cold backups. Looking
good so far but we've not actually written the code yet.


Elastic Map Reduce:

I've used it for a few one-off analytics jobs. It's incredibly nice if
you have a one-off map-reduce job you want to run over some data and you
don't want to handle all the issues with setting up or maintaining your
own haddop cluster.


Lots of the "cool" kids featured on high scalability use AWS btw - scan
through "Real Life Architectures" for stories from the trenches:

http://highscalability.com/blog/category/example

Tim