Thread by @QuinnyPig

I realized I've never done a tweetstorm on how to save money on your @awscloud bill. This will not be comprehensive in the least, it's nuanced and site specific. Here we go.

For some reason people go through their bills alphabetically. Radical idea: start with the big numbers first.

Something like 60% of global spend goes to EC2. The biggest cost savings there are "turning things off that you're not using." Then spot instances. Then reserved instances.

One year RIs hit break even in 7 months (usually), so if you don't have concrete plans to move regions / off of EC2 by June 1 make the purchase today. Your schedules will slip; it's what they do.

There's a line item buried under EC2 on your bill called Managed NAT Gateway Data Processing. That can be a giant number. With remarkably little work it can become zero. It's in addition to the data transfer cost.

Data transfer is arcane, nuanced, and anyone who claims to fully understand all of its nuances is lying. Fight me!

But do understand that you're charged for traffic between AZs--and twice as much for transfer between regions.

The exception to this is traffic between us-east-1 (Virginia) and us-east-2( Ohio). This is billed at half price, of the same as inter-AZ.

The reason for this is that even data wants to get the hell out of Ohio.

Let's talk EBS. You should almost always be using general purpose (gp2) volumes. "But Corey!" you cry. "I need better performance!"

gp2 throughput scales with volume size to a cap of 16K. io1 can go to 64K, but almost nobody needs that. 16K or less, the answer is to make the volume bigger. There's a boundary condition around 16K, but in general you'll get the same performance for 1/3 the cost.

EBS snapshots will range from "piddly" to "that's a phone number!" You can save money here, but be cognizant that you will not be praised for destroying site durability in the name of cost savings.

There's more here about spot, autoscaling, etc--but "oh, just refactor your app to be able to autoscale" is naive, condescending, and not nearly as easy as it sounds for many things. If it makes sense, do it--but it's not as easy as it may seem.

Next let's talk about S3. You can save money here if you know your usage patterns. S3 analytics help you discover what you're doing, but they take ~90 days to return recommendations.

There's an old tier of S3 called "reduced redundancy." Worse durability than standard, it doesn't get price cuts anymore so it's more expensive in most regions than good storage, and AWS doesn't talk about it anymore. Stop using it.

Infrequent Access costs less to store data, but charges you more to access it. If baby seals get more hits than your older data but you still need it online, use this tier.

There's a still cheaper option: Infrequent Access One Zone. Single AZ rather than multiple, so if AWS loses a whole AZ your data goes with it. This has never happened, but it's worth considering in a context of business risk. Usually "data that can be reconstructed."

Glacier is fascinating for long term storage. Lots of people have stored data there, few have ever restored anything from it. Originally restore charges were opaque and enormous. That is no longer true.

Glacier Deep Archive was announced at re:Invent but has not been made GA yet. That'll save you yet more. Ask me about this post-launch.

DynamoDB is fun. On-demand capacity is a win, but it's significantly more expensive; if you pay for 100 read/write units reserved, it costs the same as if you use on-demand usage of ~14 units. Plan accordingly.

RDS is its own giant billing issue, but it's highly contextual and not very interesting. I'll take a pass on it for this thread.

Lambda bills are fun to optimize except for the part where the money absolutely doesn't matter for any sane architectures. (Lambda@Edge not included.) I could optimize my Lambda costs myself, but given that I'm spending something on the order of 30¢ a month for it...

And lastly, despite what you probably believe, your company doesn't actually care about saving money on the AWS bill--they just think they do. What they really want is to understand and predict it accurately.

I will now take questions from the audience.

Oh, and if you care about this sort of thing / have a bill that's spiraling out of control, let's chat. I'm at quinnadvisory.com

One last update: I somehow didn’t realize that the existence of Enterprise Agreements is public, so I can disclose that I help companies negotiate them.