Monday, December 22, 2014

... on xen

A few years ago I was introduced to the erlangonxen project. After the initial impression they went back to the drawing board and produced a new and impressive demo. The second demo showed tens of thousands of application instances being spun up and torn down.

In the interim I introduced myself to docker and was able to produce better numbers. EOX was deploying a new environment every 300ms and I was able to deploy a docker container in under 40ms.

As I started to dig into EOX I determined that there was no real OS underneath it and that there were severe limits to the filesystem. Something like 500 files. I still think there are some limits to the design but they may not be as important as I first thought.  First of all it is significant that the application can be launched on demand and it can be readonly preventing any number of potential security problems. One can also address the multi-tenant issue by just building the front-end on demand per tenant.

Now add elixir on xen and MirageOS and things are getting hot. At the moment development is typically taking place on dedicated bare metal and companies are moving to virtual (public and private) systems like VMware, OpenStack, RackSpace, Azure, GCE and so on. And with the sudden success of Docker the entire container ecosystem is getting a much needed boost. While we seem to be moving to containers the question seems to be what is next? And as I think about it I'm starting to fall into the unikernels like MirageOS.

The fact is that the hypervisor manufacturers are doing a great job of abstracting the hardware. And many of the tools makers are starting to take advantage of that standardization. Of course a lot of people seem to have forgotten that in a bare metal world DOS was a unikernel and so this new stuff is actually old.

What makes it interesting is the density.  A moderate sized EOX application could be a few megabytes, however, an equivalent Docker base container could have a much bigger footprint. Even the Scratch and BusyBox images have their limits.

All in all, however, I'm watching the state of things. I could find myself moving back to erlang or haskell as the latest graduating class of programmers pollute go and new breed of system languages that have decades of actual experience behind them.

** don't even talk to me about generics

No comments:

Post a Comment

prod, staging, QA, dev in your CI/CD?

I've been developing with CI/CD since before it was a straw, let alone a pipeline. No, graduates of 2020 you did not design or discover ...