(Part 3 of 4 part series)
In previous articles (part 1 and part 2) I’ve emphasised Disaster Recovery (DR) design principle is simply about eliminating single point of failure for data centre, and to provide adequate service and application resilience that’s fit for purpose. Over-engineered gold plated architecture solution does not always fit the bill and conversely low-tech simple and cost effective solution doesn’t necessary mean it’s sub-standard. There are 3 common DR patterns that you are likely to find in your organisation and they are known as “Active-Active”, “Active-Passive” and “Active-Cold”. As a DR solution architect you have been tasked to implement the most cost effective and satisfactory DR solution for your stakeholders. You might wonder where to begin, Pros and Cons of each DR pattern and what are the gotchas? Well, let me tell you there is no perfect solution or “one-size-fit-all” silver bullet. But don’t feel despair as I will be sharing with you some of the key design consideration and relevant technology that is instrumental to successful DR implementation.
Network and Distance Consideration
Imagine your two data centres that are geographically dispersed, the underlying network infrastructure (e.g DWDM or MPLS) is the very bloodline that interconnects every service together such as HTTP server, database, storage, Active Directory, backup etc. So without doubt network performance and capability is rated high on my checklist. How do we measure and attain good network performance? First of all you’d need to understand the two key measurements; network latency and bandwidth and I will briefly explain them below.
Network latency is defined as the time it takes to transfer a data packet from destination A to B and expressed in Millisecond (ms). In some cases latency also includes the data packet roundtrip with acknowledgement (ACK). Network bandwidth is the maximum data transfer rate between destination A and B (aka network throughput), and the transfer rate is expressed in Megabits per second (Mbps). Both of these metrics are governed by the law of physics (i.e. speed of light) so the distance in which separated the two data centres plays a pivotal role in determining the network performance and ultimately the effectiveness of DR implementation.
Having data centres located in Sydney and Melbourne sounds like a good risk mitigation strategy until you are confronted with the “Zero RPO” dilemma. How could you keep data in-sync between 2 data centres stretched over 800Km, leveraging the existing SAN storage based replication technology, without causing noticeable degradation to storage performance? How about the inconsistent user experience being felt by users who are farther away from the data centre? Remember the law of physics? Unless you own a telephony company or unlimited funds, trying to implement synchronous data replication over long distance, regardless whether it is host or storage based replication technology, will surely cost a large sum of money and not to mention the adverse IO performance impact.
For those brave souls who are game enough to implement dual site Active-Active extended Oracle RAC cluster, the maximum recommended distance between 2 sites is 100Km. However after taking into consideration of super-low network latency requirement and relatively high cost, it’s more palatable to implement extended Oracle RAC cluster in data centres that are 10-15Km apart. You may find similar distance constraint exists for other high availability DR technology. Active-Active pattern is especially sensitive to network latency because of the constant chit-chatting between services at both sites. If the distance between 2 data centres is becoming the major impediment for implementing Active-Active DR or synchronous data replication, then you should diligently pursue alternative solutions. It’s quite possible that Active-Passive or non-zero RPO is acceptable architecture so don’t be afraid to explore all options with your stakeholders.
Mix and Match Pattern
I have come across application systems which have been architected with the flurry of mix and match DR design flair that got me slightly bemused. Let us examine a simple example. A “Category A” service (i.e. highly critical customer facing) is composed of Active-Active DR pattern for the Web Server (pretty standard), Active-Passive pattern for the Oracle database (also stock standard), and Active-Cold pattern for the Windows application server. So you may ask what is the problem if RTO is being met?
As you may recall each DR pattern comes with predefined RTO capability and prescribed technology that underpins it. By combining different DR design patterns into a single architecture will undoubtedly dilute the desired DR capability. In this example the Active-Cold pattern is the lowest common denominator as far as capability is concerned, so it will inadvertently dictate the overall DR capability. The issue being is why would you invest in a relatively high cost and complex Active-Active pattern when the end result is comparable to the lowly Active-Cold design? The return on investment has greatly diminished by including lower calibre pattern such as Active-Cold in the mix.
Another point you should consider is can the mix and match design really stand up in the real DR situation and meet the expected RTO. I have heard the argument that the chosen design works perfectly well in the isolated application DR test. What about in the real DR situation when you are facing competing human resources (e.g. Sysadm, DBA, Network dude) and system resources like IOPS, CPU, Memory, Network etc. It’s my belief that all DR design patterns should be regularly tested in simulated DR scenario with many applications, in the interest of determining the true DR capability and effectiveness. You may find the mix and match DR architecture does not work as well as expected.
Finally the technology that underpins each DR pattern could have changed and evolved over time. Software vendors often change functionality and capability with future releases so DR pattern must be engineered to be adaptive to change. As a result there’s inherited risk for mixing different DR patterns that will certainly increase the dependency and complexity for maintaining expected DR capability in the fast changing technology landscape.
Mix and match DR pattern may sound like a good practical solution and in many cases it is driven by cost optimisation. However after consideration of the associated risks and pitfalls I’d recommend choosing the pattern that is best matched for the corresponding service criticality. Although it’s not a hard and fast rule but I do find the service to DR pattern mapping guidelines below are simple to understand and follow. You may also wish to come up with different set of guidelines that are more attuned to your IT landscape and requirement.
Last but not least I’d like to bring automation into DR discussion. In the current Cloud euphoria era automation is the very DNA that defines its existence and success. Many orchestration and automation tools are readily available for building compute infrastructure, programming API and PaaS services configuration just to list a few. The same set of tools can also be applied to DR implementation with great benefits.
In my mind there is no doubt that Active-Active is the best architecture pattern, however it does come with a hefty implementation price tag and design constraints. For example some application does not support distributed processing model (i.e. XA transaction) so it can’t run in dual-site Active-Active environment. Even for the all mighty Active-Active pattern automation can further improve RTO when applied appropriately. For instance client and application workload distribution via Global DNS Service or Global Traffic Manager (GTM) needed for DR can be automated via pre-configured smart policy. Following the same idea database failover can also be automated based on well tested configurable rules. This is where automation can simplify and vastly improve the quality of DR execution.
Same design principle applies to Active-Passive and Active-Cold DR pattern as well. Automation is the secret source for quality DR implementation. Consider incorporating automation to all service components where possible. But here is the reality check. Implementing automation is not trivial and it is especially difficult for service component that is not well documented or designed, or lack of the suitable automation tools. Furthermore it is not advised to automate DR process if there is no suitable production like environment (e.g. cross-site infrastructure) to conduct quality assurance test. The implementation work itself can be extremely frustrating because you’d need to delicately negotiate and cooperate with different departments and third-party vendors. Having that said I believe the benefits are far outweighed the pain in most cases. I have known one case where automation has reduced DR failover time from 4 hours down to 30 minutes. No pain no gain right?
For those who are DevOps savvy techies there are many orchestration tools out in the marketplace that you can pick to develop the automation framework of your choice. Chef, Puppet, Jenkins for orchestration and Python, Powershell, and C Shell for scripting just to name a few. If you don’t want to build your owner automation framework then you might want to consider vendor software like Selenium, Ansible Tower or Blueprism.
In conclusion a successful DR implementation should be planned with detailed impact assessment of network latency between data centres, carefully consider the most appropriate DR patterns and relevant technology for the targeted service application, and leverage automation infused with artificial intelligence (i.e. policy or rule based) to replace manual tasks where feasible. In the next article I will be exploring the various DR scenarios presented for Cloud deployment.
This article is a guest post by Tommy Tang (https://www.linkedin.com/in/tangtommy/). Tommy is a well rounded and knowledgeable Cloud Evangelist with over 25+ years IT experience covering many industries like Telco, Finance, Banking and Government agencies in Australia. He is currently focusing on the Cloud phenomena and ways to best advise customers on their often confused and thorny Cloud journey.