Hi,

While discussing an architecture for a network service that should have high availability, I have been proposed to build a redundant system with two nodes and a load balancer in front (all three machines being run from the virtualization environment).
The purpose here is to ensure that the end users are almost always able to get served.

My feeling is that in an virtual environment that is useless. I mean, normally the redundancy and the high availability are provided by the hosts sustaining the virtual environment.
For the network service, one virtual machine allowed to run is enough to ensure the end users the same availability.

I am curious to hear from other users and LQ how they are handling similar cases.

Assuming you are talking about vmware/hyperv/xen... with a single machine acting as a load balancer, you still have a single point of failure.

Not necessarily. Hardware still fails, and software still has bugs... I run several websphere portal servers as two node clusters. Each node is front ended with IHS (think Apache with the IBM brush). We used keepalived, with amazing results.

Provided the servers are on different host blades, we are pretty well protected. Also, during certain circumstances, we can to live patching, but failing over to each node.

footnote:
We actually use keepalived for loads of different apps, I just chose the first that sprang to mind.

HTH.

If those blade are in the same chassis then you still have a single point of failure. It really comes down to how comfortable you are with your configuration.

Hi,

I thank you for your replies.

Let's break it in software and hardware. For software the two nodes will have similar if not identical sotfware and configuration. If there is bug or a misconfiguration it will be replicated on both nodes. So if one fails, chances are high that the other will fail on the same conditions. I am not really concerned about this. Also patching should not be a big concern justifying two nodes.

It is for hardware failure that are more concerned. But then I will prefer to totally rely on multiples hosts to minimise hardware failure of any of them.

So at this stage, I still think it will be a waste of resource to have load balancer and two nodes, for just one very important and high available network service.

Agreed, but the chance of total chassis failure are _very_ small. Here, we run each datacenters off two HP C7000s, so don't have to worry too much on that front. You could get the same redundancy from to separate rack mounted servers.

In the past I've used multiple 2U servers (HP DL360's) each with Citrix XenServer with (usually) 4 guest VMs

The N "app" servers were front-ended by two (light) VMs running keepalived and haproxy with these two instances lb001 and lb002 running on different physical boxes.

Web servers were handled similarly in terms of VMs for the actual servers but had Citrix Netscaler virtual appliances in front doing the balancing.

So to answer the question "Does it make sense to have a load balancer with two nodes in virtual environment?" then yes, it does, assuming you've mitigated other single points of failure.

However, from what you've described it seems you're looking for HA rather than LB, so you could just use keepalived to use a single IP that'll automatically fail-over between the two nodes.

1 members found this post helpful.

Let's say you are load balancing a load among 2 VM nodes with 4 GB of RAM each.
Why not just use one VM with 8 GB of RAM?

Because it's seldom about just resources such as RAM and threads.

Go read up on the advantages/disadvantages of vertical and horizontal scaling.

Am I reading this correctly? You think it is a waste for something that is very important and high available network service? This is exactly what this is meant for.

Anything that is of importance should be load balanced or at the least fail-over ready.

Yes, you did.
The high availability will be provided by the virtualization hosts.
Their intention is to give me a load balancer and two nodes, all hosted on the same virtualized environment.

I suggest that you forget about the word, "virtualized," and consider this to be a simple question of horizontal vs. vertical vs. combined scaling.

Consider the real hardware: virtualization is just an illusion. If the box breaks, it will take down every VM that's on it. (Usually, of course, it's a disk drive that actually breaks, so solid-state drives are probably a good idea, etc.) The presence or absence of virtualization has no impact on the physical devices that might break down, but it does give flexibility at an acceptable performance cost.

Some devices, like the load balancer itself, are unlikely to break.