Decoding AWS Reserved Instances | Hacker News Reader

Decoding AWS Reserved Instances

blog.symops.com

47 pointsabuggia23 comments

This year some of my RIs became useless to me with a chunk of the 3-year term remaining, due to the EC2-Classic shutdown. (The instance types were a couple of generations old, and no longer available in the same availability zones in the VPC network.) Since understandably no one on the marketplace wanted to buy them, I asked AWS to credit me for the unused time on the RIs, with a promise that I'd spend the credit on new RIs on the latest generation hardware. They kept saying no -- RIs are a commitment, even if they become worthless due to their own mandates, apparently. I spent 40 days on a support ticket in which they'd go silent for a week at a time and then come back with an offer for a credit conditioned on me committing to buy way more RIs than I need. Finally they just bought my unused RIs from me on the marketplace at their full value. An acceptable outcome, but I should never have had to fight AWS for this credit.

No familiar with AWS. So you got Reserved Instance Credis that are only able to specific Instances, but the Instances were no longer available to “buy“.

If that, it is insane that you have to fight that they give you your money back.

You should talk to your account manager but for non ec2 services we have been successful with exchanges by explicitly asking for an exchange for a similar/larger amount.

We had to buy the new reservations first and would then could cancel the old ones

It probably would have been better if AWS had called them "reservations" rather than "reserved instances".

NB they would never call it a lease or rental because those terms have legal repercussions in some jurisdictions.

A reserved instance is simply pre-paying for a specific amount of EC2. AWS is able to discount this because: 1. You're pre-paying 2. You might not use all the resources you're entitled to, if you don't keep a running EC2 instance "filling" the reserved instance slot. 3. You're helping them with capacity planning and capacity management; there's no guesswork involved with your workload.

Disclosure: I used to work on GCE.

We got to go second, but I prefer the naming and separation of concerns we came up with (obviously): Reservations are capacity you're holding, while Commitments are things that get you discounts.

Reservations are things your operations team uses to guarantee capacity for scaling up, doing rolling restarts, and whatever else. Commitments (and Committed Use Discounts) are primarily what your Finance teams work with.

The original Reserved Instance product combined both of these. AWS now has the same decoupling with Savings Plans and Reservations as related, separate products.

I’m not familiar with GCP but those names/concepts are easy to understand and I like the separation of concepts.

It is great to see articles like this, as Reserved Instances are a fantastic way to save some money, and they are widely misunderstood.

I work for a company [0] that specialises in Cloud FinOps, and I helped write our in-house billing system, which has to deal with _all_ of the intricacies of Reserved Instances, so I know very well how deep this rabbit hole goes.

This article opens by categorising RIs as "one of the most complex but crucial AWS offerings", but I think it perhaps glosses over some of that complexity, which might cause people to get burnt. For example, it suggests that you can safely sell your way out of trouble on the RI Marketplace, but in many cases you can't: there are 5 products you can buy RIs for (EC2, RDS, ElastiCache, OpenSearch, and Redshift), but the RI Marketplace only lets you sell EC2 RIs, and only Standard EC2 RIs at that, not Convertible.

I don't want to take anything away from this article, nor do I want to frighten people away from buying RIs, but please take care. If you aren't sure, I encourage you to use one of the many companies operating in this space to help you out (and you certainly don't have to give away 20% of your savings to do that).

0: https://strategic-blue.com

Cool to see Vantage mentioned here! I work for them, mainly on the ec2instances.info site.

https://docs.vantage.sh/autopilot/ has many details on how our RI management program works. We recently added Autopilot Controls which lets you specify categories of instances (like c5) or regions that you want managed.

One note on buybacks: they are usually capped. I believe most providers cap their "guarantee" around $20K.

I didn't know this page existed, but nice work. I like how you display this data and make is so easy to do comparisons. This is a great resource.

This article talks about the financial pros and cons of reserved instances, but not the important operational benefit:

You will always get an instance when you need it, and never get an error about not having enough instances of a certain type.

This is critical if you are using autoscaling, especially during an outage. If you are shifting your workload from one region to another at the same time as everyone else, if you have reservations and they don't, you'll get the instances before anyone else.

It breaks the whole "pay as you go" cloud model, but as long as you can find batch jobs to fill those reserved instances 1/3 of the time, you'll still come out ahead.

That used to be true, when AWS only sold AZ-specific RIs, but for a long time now AWS have (by default) sold region-level RIs instead, and these do not include a capacity reservation. You can now purchase Capacity Reservations [0] on their own.

https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-capa...

That actually depends.

Reserved instances have one form with zero operation benefits: Regional Reserved Instances. These are 100% purely billing constructs.

They have zonal Reserved Instances which do reserve capacity.

Finally they have pure capacity reservations. These cost full price. Esentially you are paying for an instance even if it is not running, but it does guarentee it will be there if you need it. but it also has no comittments.

However combining the capacity reservation with a savings plan allows you to basically recreate the net effect of a zonal Reserved Instance but with additional flexibility.

Regional Reserved Instance's discount will also apply to reserved capacity if there were not enough actual instances to cover the Reserved Instances.

> However combining the capacity reservation with a savings plan allows you to basically recreate the net effect of a zonal Reserved Instance but with additional flexibility.

This is something that has never crossed my mind. TIL.

Azure's Reservations are far superior. Whether you pay up front or monthly doesn't change the cost (there's no "interest") and you can cancel the reservation early without this "marketplace" nonsense.

It makes Reservations a no-brainer on Azure. The pitfalls with AWS make it a little more murky.

> Businesses can sell RIs they’ve purchased, setting their own price for the remainder of the contract.

You can't sell convertible RIs, I thought? Would love to find out that's wrong.

From what I've seen RIs are a trap, suitable to hardly anyone, especially not startups. If your infra is elastic or evolving at all, be really careful before going for RIs. And I would never go for more than a one year reserve.

(maybe if you are deploying K8s over RIs it's fine, you'll be able to utilise at least partially even if your infra changes - but if you have workload-per-instance or service-per-instance then be careful)

You are correct, you cannot sell Convertible RIs on the RI Marketplace.

I think it is going too far to say that RIs are a trap: the rules are pretty transparent. I agree with you that if your infra is elastic or evolving then you need to be very cautious about making large or long-term commitments.

I think the convertibles are a bit of a trap, though - even though the rules are transparent. You might overprovision to begin with (which you may not realise), then you can't even sell them.

Almost everyone has some kind of base workload, whether it’s RDS or EC2 operating some specific piece of infra.

Some people may not be able to RI their entire workloads, but EVERYONE can reduce their AWS spend partially using RIs. It may only be a little bit but better in your pocket than Amazons.

Reserved instances are effectively a financial construct.

A lot of human hours go into managing them, buying them, selling them, making sure they are efficiently used, etc.

I do wonder if the human effort invested in such things is actually a net gain over a fixed price. Clearly there is gain for the companies using RI's, but is there gain for the economy as a whole?

> Reserved instances are effectively a financial construct.

Almost true. It's also a powerful lock-in mechanism.

> Reserved Instances are a major piece of AWS Billing, accounting for $100B in revenue along side Saving Plans (SPs).

I think I know what you mean, but this reads like Reserved Instances earn AWS $100 billion/year which is almost impossible. I don't even think AWS earns that much annually.

Great to see https://www.usage.ai/ mentioned here. I'm the founder & CEO, happy to answer any questions here.

This is written from an interesting angle and I'm not sure who it is meant to serve. It "and then's" Savings Plans but for the majority Saving Plans replace RI and are infinitely more flexible and easy to use.

TLDR if you are choosing RI's over Saving Plans the basic info conveyed in this article is probably already know if you don't you should probably just be choosing saving plans.

Reply on news.ycombinator.com