AWS Reserved Instances, Part 1: Theory

AWS's Reserved Instances began in 2009 as a simple means of obtaining a discount through commitment. Today, the pricing model still gives end users the opportunity to make substantial discounts of more than 50%, but they are not as simple as they once were. AWS has, quite correctly, responded to customer feedback, and has made Reserved Instances more flexible with a greater range of options, such as the ability to convert instances and apply them to different zones in the same region. But, inevitably, greater choice brings new challenges to end users in making the right decision. In this first of two reports, we explain some of AWS's additions to the scheme. In Part 2, we put theory into practice and derive a breakeven analysis with recommendations.

Reserved Instances have many benefits. The Cloud Price Index average discount in the US is about 30% – for AWS, discounts of over 60% can be achieved. For enterprises that like predictability, the option to commit to a monthly fee or to pay up front can be attractive, and for those with mission-critical requirements, capacity guarantees can be attractive. But with more choice, end users need to take more time to fully assess which option is the best. AWS and third parties have tools that can help, but Google has taken a more proactive approach to cost-optimizing – its sustained-use pricing model is fully automatic and rewards end users with cuts of up to 30%.

AWS EC2 customers can purchase virtual machines in a large range of specified sizes using three models: on-demand pricing, paid for by the hour; spot pricing, unguaranteed access based on a bid in a form of market auction; and Reserved Instances (RI) – the focus of this report. Reserved Instances are a billing concept – they aren't attached to a specific resource, and can be applied to resources as and when needs dictate. Think of this as prepaying for a soda, for which you are given a cup. The receipt states that you are entitled to a soda on this visit, but doesn't say which specific soft drink – you just apply the purchase to whatever you fancy. Similarly, each RI has a set of parameters that dictates when and where it can be applied, such as location and size of instance. Originally just for EC2, the model is now available for databases via RDS.

End users can pay for a full term up front, commit to paying a monthly fee or a combination of both (named All Upfront, No Upfront RI and Partial Upfront, respectively). In general, percentage savings increase with the length of term (one- or three-year terms are offered) and the size (and therefore on-demand price) of the virtual machine. As expected, consumers that pay up front are rewarded with a bigger discount for that virtual machine, and those that pay on a monthly basis are rewarded with the smallest discount. End users pay for the whole term regardless of how much is consumed.

Although cost savings are a big attraction for end users, capacity and predictability are also central to enterprise requirements. A Reserved Instance is linked to a specific availability zone where capacity is guaranteed. If there is a spike in demand across the zone, RIs are given priority. If an end user is operating a mission-critical application based on on-demand resources, it might not be able to scale when the time comes. With RIs, the capacity should always be there. A further attraction we hear from enterprises, in particular, is predictability. Few enterprises are built to be cloud-native; there are still fixed budgets, processes and controls in place. A cloud application that can burst when needed might be attractive to marketers and developers, but to the CFO holding the purse strings, it's a different matter. Many enterprises still like the old-fashioned, predictable and low-risk model of knowing expenditure before it happens.

What's in it for AWS? Cloud providers naturally want to capture capital up front to invest in infrastructure and reduce the cost of capital, so it makes sense to give those that can pay the full amount up front the biggest discount. A monthly commitment is also attractive to providers (it brings in regular cash flow), but there is increased risk on the part of the providers – they may invest in infrastructure and give the discount only for the consumer to default on their commitment. For standard three-year terms, AWS doesn't offer a 'no up-front option,' probably due to this increased risk. AWS also benefits from predictability – data on RI purchases can be used to aid capacity planning, which can reduce costs by lowering wastage while optimizing revenue by having capacity available when needed.

RI Marketplace

RIs define a size of virtual machine in a specific location, and originally couldn't be applied to differently sized virtual machines. This rigidity was unattractive to some end users that wanted to change their architectures to meet changing needs. To some, a three-year commitment on such rigid terms was off-putting. In 2012 AWS launched its RI Marketplace, which allows end users to sell partially consumed RIs to other users. In addition to letting end users recover some capital to invest in differently sized instances, this allowed users to buy and sell RIs for a range of terms, not just the one- and three-year terms offered by AWS.

The marketplace significantly improved the proposition of RIs. Although selling RIs was not guaranteed, it did mean end users weren't completely trapped once a commitment to an RI had been made – we believe its key feature is reassuring users that they can move. AWS benefitted by selling more RIs, thereby improving cash flow, capacity planning and commitment as mentioned previously. It also benefits from a 12% admin fee.

Modifications

In 2013 end users were given the ability to modify RIs between zones within the same region, and between standard and Virtual Private Cloud models. A few months later, end users were given the ability to combine or split RIs to support smaller or larger virtual machine sizes. These 'modifications' are always done within an instance family (m1, m2, m3 or c1), and are always expressed in terms of normalized units. So to make a large instance with 'normalization factor 4,' one needs to combine two medium instances with 'normalization factor 2.' The benefit of this flexibility is apparent, but this added another consideration to end users making procurement choices.

In 2014 AWS released better reporting on utilization of reserved instances to allow end users to make better decisions on placement and usage. Prior to late 2014, AWS used a slightly different RI model, which we won't confuse matters here by explaining, but more details can be found in this report.

Scheduled Reserved Instances

In January AWS announced scheduled RIs that can be purchased by customers to operate at specific times. The ideal workloads for these RIs include batch-processing applications, such as overnight video rendering, data processing and archiving. Scheduled RIs give end users assurance of capacity during critical times with an automatic discount, without having to achieve a heavy utilization capacity.

Regional flexibility

In September AWS announced that it would allow end users to use their reserved instances in different regions from where they were attached. However, end users give up a guarantee of capacity in the original availability zone by doing this. This is still an attractive proposition – if demand suddenly shifts to a different part of the world, RIs can just be moved overseas for cheaper resources.

Convertible Instances

Also in September, AWS announced a different class of RI – Convertible RIs – that allows end users to change the instance type, operating system or tenancy (standard or virtual private cloud) without resetting the term. However, the end user might need to pay an upgrade fee to do this, which represents the difference in price between the current RI value and the future value. AWS states that most end users are unlikely to convert instances much (if at all) throughout the term, but Convertible RIs provide an 'insurance policy' for those end users that want the reassurance that they can change when needed.

In Part 2 of this series, we perform a breakeven analysis of a Reserved Instance and provide guidance for those looking to gain cost benefits.