Pure Storage ActiveCluster – Background Information

I’ve been doing a bunch of research into Pure Storage’s ActiveCluster product recently. I was all set to do an article that explains how to set it up and what a vSphere Metro Cluster looks like with it in place, but Cody Hosterman has beaten me to the punch. Given that it’s more his job than mine to write this stuff, and that he works for Pure Storage, I’m okay with that. In any case, I thought it would be worthwhile to jot down some thoughts and notes and share some links to Cody’s work, if for no other reason than it gives me an aggregation point for my thoughts.

 

Introduction

I was lucky enough to be at Pure//Accelerate in 2017 when ActiveCluster was announced and covered it at a high level here. If you’re unfamiliar with ActiveCluster, it’s “a fully symmetric active/active bidirectional replication solution that provides synchronous replication for RPO zero and automatic transparent failover for RTO zero. ActiveCluster spans multiple sites enabling clustered arrays and clustered ESXi hosts to be used to deploy flexible active/active datacenter configurations.” (https://kb.vmware.com/s/article/51656).

[image courtesy of Pure Storage]

 

Components

There are a few bits that are needed to make ActiveCluster work (besides Purity 5.0 on your FlashArray):

  • Replication Network;
  • Pods; and
  • Pure1 Cloud Mediator.

 

Replication Network

The replication network is used for the initial asynchronous transfer of data to stretch a pod, to synchronously transfer data and configuration information between arrays, and to resynchronise a pod. For this network to work, you should note the following criteria apply:

  • The maximum tolerable RTT is 5ms between clustered FlashArrays;
  • 4x 10GbE replication ports per array (two per controller). Two replication ports per controller are required to ensure redundant access from the primary controller to the other array;
  • 4x dedicated replication IP addresses per array;
  • A redundant, switched replication network. Direct connection of FlashArrays for replication is not supported; and
  • Adequate bandwidth between arrays to support bi-directional synchronous writes and bandwidth for resynchronizing. This depends on the write rate of the hosts at both sites.

So, you need to know (and understand) your workload, and you need some reasonable bandwidth between the arrays. This shouldn’t be unexpected, but it’s clearly well suited to a metro deployment.

 

Pods

A Pod is a replication namespace. Once a pod is created, the pod (and the volumes inside it) can be controlled from either FlashArray. If you create a snapshot, that snapshot is created on both sides. If snapshots exist on the volume before it’s added to the pod, those snapshots will be copied over when you add it in. The pod itself acts as a consistency group.

 

Pure1 Cloud Mediator

The Pure1 Cloud Mediator is used to arbitrate split-brain scenarios. It sits in the cloud and keeps an eye on stuff. Think of it as the Vanilla Ice of the ActiveCluster (before he went off and did moto-x and renovation shows). For “dark” sites, an on-premises mediator (VM) can also be deployed.

 

A Few Other Notes

A few other things to note about the behaviour of ActiveCluster:

  • Data reduction is performed independently between arrays. This is cool because you might have a mix of workloads at each data centre;
  • If the arrays lose connection to the mediator they will continue to serve data and synchronously replicate as long as array to array communication is active; and
  • If both arrays lose communication with each other and with the mediator, this is a dual failure and both the mirrored volumes become unavailable until communication with the other array or the mediator can be re-established. Non-mirrored volumes would not be affected in this instance and would still be accessible.

 

Disaster Avoidance Or Recovery?

Before deploying ActiveCluster, you should think about what kind of goal you’re trying to achieve. Disaster Avoidance assumes that some element of the primary site (Site A) is unavailable due to a disaster. DA uses synchronous replication only and requires a stretched cluster technology (such as VMware vSphere Metro Cluster) to provide active / active workload availability access both sites. Disaster Recovery, on the other hand, assumes that workloads are deployed in an active / passive configuration across sites. There are advantages to each approach, depending on what your recovery point objective (RPO) is, and what your recovery time objective (RTO) is. If you have a very low RPO and RTO requirement, the added expense of deploying a synchronous replication solution (not the Pure bit, but the supporting infrastructure) is worth it. If you have a greater tolerance for a higher RPO and / or RTO, an asynchronous solution (and the less stringent replication network requirements) may be a better fit for you.

You should also think about whether the topology you’re deploying is Uniform or Non-Uniform. A Uniform configuration provides hosts with access across Sites. This requires a bit more investment in terms of stretched FC fabrics (assuming you’re using FC and not iSCSI). This is generally the topology deployed for metro clusters.

You might decide, however, to deploy a Non-Uniform configuration for simpler disaster recovery. In that case, there’s no requirement to have cross-site FC links in place, but your time to recover will be impacted. You’ll also want to look at something like VMware Site Recovery Manager to orchestrate the recovery of workloads at the secondary site.

 

Conclusion

Whilst I think ActiveCluster is a very neat piece of technology, you should be doing a whole lot of thinking about other (possibly very boring) stuff before you take the plunge and decide to deploy vMSC sitting on an ActiveCluster environment. Disaster Avoidance (and Recovery) require a lot of planning and understanding of what’s important to your business before you deploy a solution. In the next little while I hope to be able to report back with some results from testing, and talk a bit about other protection scenarios, including metro clusters with asynchronous protection off to the side.