Introduction

Let me take you on a journey through history. In the beginning, there was the tape loader. You loaded the magnetic tape into the case, pressed play and waited 20 to 40 minutes with your finger hovering over the stop button waiting for the game to start. That was a lot of fun! (not).

When I first started working in IT, I was doing a job not dissimilar to what I do nowadays. Back then servers were expensive and it took a long time to get them up and running. The quickest I could get a new service deployed was roughly two weeks. This involved purchasing the servers, waiting for them to be delivered, building them, taking them to the data centre to be racked and installed, setting up the Cisco routers, storage, etc. There was a long wait before you could enjoy the service, like playing your favourite 8-bit game.

Then Virtualization came along. Xen and VMware were the first incarnations I worked with. It drastically reduced the number of trips to datacentres as we just needed to increase capacity from time to time as opposed to buying new servers for every single project. No more time waiting.

But a few years later, cloud was the buzzword.

We don’t need a data centre anymore, we can put everything into the cloud!

The advent of cloud computing

Although cloud computing has been around for longer than most people think (AWS launched in 2006!) it didn’t become popular until a few years later. They promised unlimited resources at your fingertips, instant availability and cheaper than running your own data centre.

This became very attractive for both engineers and managers. For us, it was all about how quickly we could get new services deployed. Just a few clicks and we’re up and running. Need more resources? No sweat, I’ll add a few more servers.

For management, it was about price and time to market. I worked for fast-paced industries where beating the competition was a high-stakes game. You needed to be faster than them!

Also, you ought to remember there is more than just computer people. Managing a data centre is team’s work: site management, access control, networks engineers, storage engineers, etc.

You start to see why the cloud became popular with management. Fewer people are required to manage a larger number of servers and you can deliver the services at a fraction of the time and at a promised cheaper price.

Why are some companies are going back to on-premises then?

Price

This is probably the number one complaint I heard from friends and customers. I was having a coffee with a friend of mine a few weeks ago. He’s an IT manager for a large media company in the UK. He was telling me how the company started moving services from local data centres to the cloud a few years ago but they are now going into reverse.

As an example, he told me their Apache Kafka servers had a 50% increase in performance and a 70% reduction in price. I have no way of verifying this is true but I do trust my friend and it’s not dissimilar to what we have been seeing.

Last year I was honoured to speak at ApacheCon North America. My colleague Hayato and I did a lot of research on how the cloud providers' disk performance affected the running of Apache Cassandra. Our conclusions were quite daunting: cloud is slow and expensive for data-intensive applications.

Performance

As I hinted in the previous point, you will always get a different performance in the cloud than bare metal. You are sharing resources with many other people and will never get full use of a CPU.

Disk performance can be problematic for data-intensive applications. At Digitalis we manage many databases such as Apache Cassandra and Postgres and others such as Kafka and Pulsar where disks must be fast if you don’t want it becoming the bottleneck of your applications.

During our research for ApacheCon we found local SSD disks to be easily twice as fast as the best cloud provider disk at a fraction of the price. And whilst some cloud providers will offer you local disks (ephemeral storage) it is often impractical due to the high level of failures. Only last week I had to deal with 6 servers using ephemeral storage dying which needed to be rebuilt.

Meeting halfway: hybrid cloud

In the last couple of years, I went full circle. I used to be an advocate of cloud only but no longer. There is definitely a place for bare metal. While it is a bit more involved and not as quick to deliver as full virtual, there is also much to gain.

You don’t need to manage your own data centre, you can easily lease/rent servers from hybrid cloud providers to enjoy the benefits of bare metal.

The idea behind using virtual machines was that you could use the spare capacity of your hardware to run multiple applications. But this was all before containerized applications were all the rage.

Does it make sense any more to run Kubernetes on virtual machines? Yes, auto scaling is fantastic but not really required by all companies and workloads. If you have predictable growth, surely it’s just about planning when to add more nodes in advance.

What about data applications? Databases benefit hugely from running in bare metal where they can get access to all resources offered. And the local SSD storage is going to be much cheaper and faster.

This is why I think a hybrid approach is the best. You can run data-intensive and containerised applications in bare metal and use virtual machines for small deployments that cannot be run in Kubernetes and burst workloads when the demand outstrips the available resources.

Conclusion

Cloud is great. It is easy. The time to market is very small. Just some ansible or terraform and you’re up and running in no time. But it is expensive, especially for running data applications.

Don’t compromise. Do your own research, do the maths and weigh all the options. There are good cases for running everything in the cloud, everything in your own data centre and for sure, a hybrid approach. If you don’t know which one to choose, ask us.

‍