These days we often have to run a large number of servers, and the times where we could afford to manually log into each one to do system administration tasks are mostly over.
It turns out that there are always different approaches to deal with this. In most cases we are talking about virtual hosts, so we have a layer between via the visualization that can help us. We can have a number of master images and create virtual images from those even on demand in a matter of a minute. In case of MS-Windows it is an issue that they have some internal UUID as host-id which should be unique and which is heavily spread throughout the image, but this issue can be ignored if we do not worry about windows domains. Usually we do and I leave dealing with these issues to MS-Windows-experts.
Talking about Linux, we only need to make sure that the network interface is unique, which it is if we use hardware and do not mess around with it, but it is not necessarily if we use visualization and virtual network devices. This issue needs to be addressed, but it is well supported by common visualization tools. Another point is the host name. This is not too hard, because we only need to change it in one or two places, which can easily be done by a script. We can mount the image and do the change. Now the image can contain a start-up script that discovers on boot that it is a fresh copy and uses its host-name to retrieve further setup from some server. And we just have to maintain there which host has which setup. These can be automated to a very high extent. Then we can for example request a certain number of servers with certain software and configuration via an web interface. This creates new host-names, stores the setup with these host-names in its setup table, creates the virtual images, deploys them on any available hardware server and once they have stared they retrieve their setup from the server. We can also have master images that already contain certain predefined setups so that this second step is only needed for minor adjustments. We have to assume that these exist. Yes, this is called cloud technology.
If we keep the data somewhere else, these servers can be discarded and new ones can be created, so there is no need to do too complex stuff on them. Of course we want to run our software on them. So the day long procedure to install our software is not attractive any more. We need mechanisms that can be automated.
Running real hardware is a bit more demanding and for larger servers that might even be justifiable, because they do a lot of work for us. Quite often it is possible to actually do mechanisms quite similar to the virtual world even on real hardware. It is possible to boot the machine from an USB-stick which copies fetches an image and copies it on disk. Again only the host-name needs to be provided and then the rest can be automated. Another approach is to initially boot via the network, which is an option that most of us rarely use, but which is supported by the hardware. For running a large server farm such a hardware and bios setting can just be initially the default and from there machines can install and reconfigure themselves. In this case we probably need to use the Ethernet-address of the network device as a key to our setup table and we need to know what Ethernet addresses are in use. It is a big deal to set up such an environment, but once it is running, it is tremendously efficient. Homogenous hardware is of course essential, maybe an small number of hardware setups, but not a new model with each delivery. It is not enough that the new hardware is named the same as the old one, it needs to be able to run the same images without manual customization. It is possible to have a small number of images, but having to supply already different images for different server setups multiplying there number with the number of the hardware setups can grow out of control, if one of the numbers or both become too large.
Now we also have ways to actually access oure servers. there have been tools to run a shell just simultaneously on n hosts to do exactly the same at once. This is fine if they are exactly the same, but this is something we need to enforce or we need to accept that servers deviate. There are tools around to deal with these issues, but it is actually quite reasonable to do a script based approach. What we do is using ssh-key-exchange to make sure that we can log into the servers from some admin server without password. We can then define a subset of the set of our servers, which can be one, a couple, a large fraction, all or all with a few exceptions, for example. Then we distribute a script with scp to all the target machines in a loop. We run this on each target machine using ssh and parse the outputs to see which have been successful and which not. Here it is usually a good idea to have a farm of test servers to try this out first and then start on a small number of servers before running it on all of them.
The big bang philosophy of applying a change twice a year on the whole server landscape is not really a good idea here, because we can loose all our servers if we make a mistake and this can be hard to recover, although still have the same tools and scripts even for that, unless we really screw things up. So in these scenarios software that supports the interfaces of the previous version for its communication partners is useful, because it allows to do a smooth migration.
Just to give you a few hints: During some coffee break I suggested that Google has around a million servers. Even though there is no hard evidence for this, because this number seems to be confidential and only known to Google employees, I would say that this is a reasonable number. For sure they cannot afford a million system administrators. The whole processes needs to be very stream-lined. Or take the hosting provider where this site is running on. It is possible to have virtual web-hosts, in this case it is multiple sites running on the same virtual or physical machine sharing the same Apache instance with just different directories attached to different URL-patterns. This is available for very little money, again suggesting that they are tremendously efficient.
