I wrote some time ago about the Sun Grid service. The prices they were quoting seemed competitive with the actual use we were getting out of our Beowulf cluster. Over the last few months, however, several groups at OCLC are experimenting with the cluster, so the in-house cluster might be ahead on cost, and at the time the Sun Grid was more promise than fact. But last week I went ahead and set up a 'virtual server' on a commercial service.
I've been trying to get a 'quick' or 'live' search demonstration up for some time now, but because of various problems with our main Research Linux machine in terms of capacity and stability, it's taken longer than I wished. Another problem is that I tend to prototype this sort of thing in Python using Python's simple built-in HTTP server. Our security people like to know what software is exposed to the outside world, which pretty much restricts us to using Apache or Apache Tomcat for HTTP. I see their point, but it does present a barrier for making something public quickly.
So, last week I decided to look into Python hosting services. Some of these basically offer Zope or Plone support, but I was attracted by the more general service of renting a virtual server. For a modest cost (looks like my costs will be in the $30-$50/month range) they give you root access to something that looks like your own server. I can run anything I want with no fear that I'm putting OCLC's servers at risk. Within a couple of hours of deciding to try it, I had an application up and running. Of course, along with this freedom came the pleasure of playing sys-admin on my virtual server, but I might be able to convince someone else to do that for me.
Looking around, there are quite a few of these services (I found some links off of Python.org). Most of them offer packages with set prices for a certain amount of disk, memory, bandwidth and cpu time, but OpenHosting.com bills on the basis of what you use. Since my current use is very high on memory use, but low on everything else, that suited me better. And if I run out of memory on this server, I can ask for another one. Support is pretty good here at OCLC, but they have a production mindset and like to schedule things in an orderly way. For experimental services, renting resources as needed might be easier. We'll see, but I call this first try a success.