Evaluating Cloud Computing Uptime SLAs
Last week's Windows Azure Storage outage made me thinking how many of us evaluate the vendor's Service Level Agreement (SLA) before they decide to deploy workloads in the cloud. I bet many think about it only when it is too late.
Let's take Windows Azure SLA and see how we as consumers of the cloud services are protected in case of downtime. Before all though I would like to point out that it is in the nature of any service (public or private) to experience outage once in a while - think about power outages that we hear about or live through every winter. It is important to understand that this will happen and as users of cloud services we need to be prepared for it. In this post I will use Windows Azure as example not because their services are better or worse than the other cloud vendors but to illustrate how the SLAs impact us and how they differ from vendor to vendor.
Each SLA (or at least the ones that bigger cloud vendors offer) contains few main sections:
- Definitions - defining the terms used in the document
- Claims - describing how and under what terms one can submit a claim for incidents as well as how much you will be credited
- Exclusions - describing in what cases the vendor is not liable for the outage
- The actual SLAs - those can be two types:
- Guaranteed performance characteristics of the service
- Uptime for the service
Looking at Windows Azure SLAs web page the first thing you will notice is that there are different SLAs for each service. You don't need to read all of them unless you utilize all of the services the vendor offer. The main point here is that you need to read the SLAs for the services you use. If, for example you use Windows Azure Storage and Windows Azure Compute you will notice that the uptime for those differ by 0.05% (Compute has uptime guarantee of 99.95% while Storage has uptime guarantee of 99.90%). Although this number is negligible at first sight using an SLA calculator you will notice that the expected downtime for Storage is twice as much as the expected downtime for Compute. It is obvious that the closer the uptime is to 100% the better the service is.
The next thing that you need to keep in mind is the timeframe for which the uptime is calculated for. In the case of Windows Azure the uptime is guaranteed on a monthly basis (for both Storage and Compute). In comparison Amazon's EC2 has annual uptime guarantee. Monthly SLA guarantees are preferable because you will avoid the case where the service experiences severe outage in particular month and stays up the rest of the year. Just to illustrate the last point imagine that EC2 experiences outage of 3h in particular month and stays up for the next 11 months. This outage is less than the 99.95% guarantee or 4:22:47.99 hours acceptable downtime per year and you will not be eligible for credit for it. On the other side if the SLA guarantee is on a monthly basis you will be eligible for the maximum credit for it because it severely exceeds the 21 minutes acceptable downtime per month.
One note about the acceptable downtime. In reality hardware in cloud data-centers fails all the time, which may result in downtime for your particular service but will not impact other services or workloads. Such outages are normally covered by the exclusion clause of the SLA and are your own responsibility. You should follow the standard architectural practices for cloud application and always make your services redundant in order to avoid this. The acceptable downtime metric is calculated for outages that impact vast amount of services or customers. Surprisingly though nowhere in the SLAs is mentioned how many customers need to be impacted in order for the vendor to report the outage. It may happen that a rack of servers in the datacenter goes down and few tens of customers are impacted for some amount of time. If you are one of those do not expect to see official statement from the cloud vendor about the outage. As a rule of thumb if the outage doesn't show up in the news you may have hard time proving that you deserve credit.
The last thing to keep in mind when evaluating SLAs from big cloud providers is the Beta and trial services. It is simple - there are no SLAs for services released in Beta functionality. You are free to use them at your own risk but don't expect any guarantees for uptime from the vendor.
When the so called secondary cloud providers are concerned you need to be much more careful. Those providers (and there are a lot of them) build their services on top of the bigger cloud vendors and thus are very dependent on the uptimes from the big guys. Hence they don't publish standard SLAs but negotiate the contracts on customer-by-customer basis. Most of the time this is based on the size of business you create for them and you can rely on good terms if you are big customer. Of course they put a lot of effort in helping you design your application for redundancy and avoid the risk of executing the SLA because of primary vendor outage. In the opposite case where you are a single developer you may end up without any guarantees for uptime from smaller cloud vendors.