::: Notes
Photo by Karolina Grabowska from Pexels https://www.pexels.com/photo/roll-of-american-dollar-banknotes-tightened-with-band-4386476/
:::
::: Notes
Foto de Andrea Piacquadio en Pexels
:::
::: Notes
Foto de Pok Rie en Pexels https://www.pexels.com/photo/dell-motherboard-and-central-processing-unit-1432675/
:::
::: Notes
Foto de RODNAE Productions en Pexels https://www.pexels.com/es-es/foto/madera-hombre-gente-mujer-7080463/
:::
| c5a.xlarge | t3a.large | |
|---|---|---|
| GB of RAM | 8 | 8 |
| CPUs | 4 | 2 |
| Number of instances | 10 | 20 |
::: Notes
Comment we will have two times more ram with the second option t3 is a bursting instance Photo by Jay Johnson from Pexels https://www.pexels.com/photo/snow-wood-light-people-6414384/
:::
::: Notes
Let's say you arrive to the desert and then the photosession starts. You enjoy it deeply, how many cool things are you going to create today.
These series of photos are made by Cottonbro, available on pexels
:::
::: Notes
After two hours under the Sun, it doesn't feel so exciting anymore
:::
::: Notes
And by the end of the session you just want to kill yourself, exhausted and thirsty
:::
::: Notes
Now a cloud provider appears and shows you a super fresh glass of lemonade and tells you: this is what you are going to pay me for being happy. That is on-demand price.
:::
::: Notes
Approximately $15K per year
:::
::: Notes
Your customer is surprised. You feel surprised, too, probably. But this is only for one cluster.
:::
::: Notes
And you have three environments, and you want to keep them on separate clusters. And it is more like $45K per year.
:::
::: Notes
No worries: someone from AWS appears and tells you: commit to one-to-three years of resource/money usage and you will get a wonderful discount.
Photo by Mikhail Nilov, on pexels
:::
::: Notes
AWS invoices can be very complex, but it is important to understand how they work to make sense of the discount options.
Photography by Elina Fairytale.
:::
::: Notes
This is a (tiny) part of the bar graph created by grouping the concepts of the invoice, and filtering by instance type.
:::
::: Notes
We decide to go with four reservations for the blue instance, two for the orange, and zero for the yellow one (as it usage is sparse).
As we can see, we don't always get a discount for all our instance/hour, as there are moments in which we have more instances that reservations.
:::
::: Notes
But that is not all: we also waste some reservations, because we don't have a regular usage of the instance (yellow case), or we have less instances of a type in a given moment that the reservations made.
:::
::: Notes
26-28% savings
:::
::: Notes
In any case, Reservations Plans provide an alternative discount method with a much more simple approach: commitment to spend money. The granularity is still per hour, so it is not as straightforward as it looks like.
Picture by Kenny Boy
:::
::: Notes
37-40% savings
Saving plans are reservations 3.0. The original reservations where focused in commitment to particular resource types. Saving plans are oriented to commitment of spent, much more flexible and logical.
With EC2 Instance Savings Plans you specify the family, for maximum discount. With AWS Compute Savings Plans you just set an amount of money that is shared between EC2, Fargate and Lambda
:::
::: Notes
It is possible to pay for the saving plan each month,or give away all the money in advance. As it can be seen in the following diagram, it doesn't make too much sense as the additional savings are small.
Photo by Sharon McCutcheon
:::
::: Notes
41-44% savings
:::
::: Notes
A woman from AWS appears to you in the middle of the ir data center and says "look, I've plenty of VMs. Borrow them for a fair price."
Photo by WinSon
:::
::: Notes
Spot instances are not related to any bidding. They are very stable, the fluctuating price changes per AZ, it is very easy to blend different types on the same autoscaling group and in general they are a very safe option for many scenarios with a hugh discount.
:::
::: Notes
Previously, it was necessary to install an agent on each node for handling spot capacity interruptions. This is not needed anymore, thanks to the ASG autoscaling capacity. In a managed group, the ASG will automatically optimize the node group for maintaining the workload capacity.
:::
::: Notes
Most workloads on Kubernetes should be stateless, unless you have taken very bad decisions in your life. That is what happens with the replica set of the lower line: one of their pods has been eliminated, but that is not a problem by design. The exception are the databases that are being run inside your cluster, because you love danger situations. You can see one of them in the photo, painted in red. Well, taints and tolerations can be used to send those pods to a nodegroup composed of regular instances, instead of spot ones.
:::
::: Notes
And remember: all of this without commitments. Free yourself from that contract. Use spot instances.
Photo by Ketut Subiyanto
:::
::: Notes
Skyscanner is deploying 100% of their capacity using spot instances, including production.
:::
::: Notes
60-72% savings
:::
::: Notes
What did your grandmother said? "Kid, turn off the lights!!!". Always follow her advice.
Photo by Edu Carvalho
:::
::: Notes
Working for 16 hours each day from Monday to Friday 81-86% savings
:::
::: Notes
Not all your clusters require the same capacity. It is pretty clear that a dev environment can be much smaller than the production one.
Photo by Elviss Railijs
:::
::: Notes
94-96% savings expecting a 30% average capacity
:::
::: Notes
Ok, we are almost done. Just keep up with your attention for a few minutes.
Photo by Andrea Piacquadio
:::
::: Notes
72-74% savings without any risk derived from spot market With full spot adoption savings goes to 78-84%!
:::
::: Notes
Uff, ok. Good work.
Photo by Zichuan Han
:::
::: Notes
73 dollars per month and cluster ($876 per year and cluster)
Photo by pixabay on pexels
:::
::: Notes
EBS: $1.2 per GB and year
Photo by Skyler Erwing on Pexels https://www.pexels.com/@skyler-ewing-266953
:::
::: Notes
Photo by Mark Glancy
:::
::: Notes
Sergio garcia: https://www.linkedin.com/in/sergiogarciagil85/ Javi Moreno: https://twitter.com/ciberado/
:::