Decoding AWS Reserved Instances(blog.symops.com) |
Decoding AWS Reserved Instances(blog.symops.com) |
If that, it is insane that you have to fight that they give you your money back.
We had to buy the new reservations first and would then could cancel the old ones
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.
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 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).
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.
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.
https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-capa...
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.
This is something that has never crossed my mind. TIL.
It makes Reservations a no-brainer on Azure. The pitfalls with AWS make it a little more murky.
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)
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.
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.
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?
Almost true. It's also a powerful lock-in mechanism.
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.
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.