← Back to Blog

Our Scanner Guessed at AWS Prices. We Made It Stop

When a price was missing, our detectors filled in a number. On 2026-11-14 we checked the tables against the AWS Price List: guesses, stale rows, a double count.

Rick Wise6 min read
Our Scanner Guessed at AWS Prices. We Made It Stop

A waste scanner's most important output is a dollar figure. If the figure is wrong, the rest of the finding does not matter. On 2026-11-14 we read our own detectors' price tables against the AWS Price List, and several of them were wrong in the same way: when the table had no price, the code made one up.

This post is the list of what we found and the one rule we ended up with.

The problem

Every detector that estimates a saving needs a price. For an RDS instance that is an hourly rate for its class. The table behind it only knows the classes someone added. When a class was missing, the code did not stop. It used a default.

Here is what that looked like:

  • RDS. db.r6g and db.r7g had no row and were priced at an invented $0.10 an hour. The Idle RDS check misstated them, and the Aurora Serverless v2 check could not fire at all.
  • EMR and EC2 previous generations. c4.xlarge sat at $0.100 an hour in our table. The list says $0.199. m3.xlarge was $0.133 against $0.266. The old types were priced below their replacements, so a "move c4 to c5" finding could never fire. A type with no row was priced at $0.20 an hour.
  • WorkSpaces. Two tables disagreed with each other, and an unknown compute type was priced at $35 a month or $0.40 an hour.
  • Redshift and OpenSearch. A node or instance type with no price used $1.00 or $0.15 an hour. The OpenSearch copy of the table disagreed with the Price List on seven types.
  • Lightsail databases. A substring match priced the xlarge and 2xlarge bundles as large: $115 instead of $245 and $490. Any unknown bundle got $15.
  • Kinesis. One shard-month was typed in as $11 in two places. The Price List rate is $0.015 per shard-hour, which is $10.95 a month. It is a small difference, but two numbers for the same thing is how larger differences start.

None of these was a typo in a single check. It was one habit repeated: a missing price was treated as a number instead of as a missing price.

What we changed

We replaced the guessed rows with list prices and wrote down where each came from. The source for the RDS rows, for example, is the public AWS Price List for Amazon RDS in us-east-1, MySQL Single-AZ On-Demand, which is the basis of every row that was already in the table. The EC2 and EMR rows come from the same list with the publication date recorded in the change.

Then we added the rule:

A resource without a known price is skipped. The scan says "price unknown, saving not estimated".

That is the whole rule. If a detector cannot price something, the finding does not carry an invented saving. You see that the price was unknown, and you can decide what to do about it.

Two smaller honesty fixes came with it:

  • Idle MSK now includes the broker storage in its saving, because deleting the cluster also stops the storage. A source that cannot read the volume size leaves storage out and says so.
  • The Step Functions check no longer says transitions were "attributed to retries". The failure share is an estimate, and the finding says that.

Two more places we showed a saving we did not have

The price tables were the biggest group, but the same audit found two others.

One EBS volume counted twice. Three checks can flag the same volume: an oversized volume, a gp2 volume that should be gp3, and one with over-provisioned IOPS. Every surface sums the savings, so a volume flagged by two of them was counted twice in the headline. Now the finding with the larger saving on a volume keeps it. The others stay visible at $0 and name the finding that carries the saving.

A saving that was a risk. The excessive-timeout check on Lambda reported a saving that was really an exposure: an assumed 1% stuck rate times the cost of a run that hangs until the timeout. That is a risk, not money you get back. The check now reports $0 and keeps the exposure in the finding's details, still shown.

Canned dashboard cards. In the hosted product, the insights service called the cost breakdown without a required argument. The call failed. The handler then returned two canned cards with potential savings of $50 and $100. We fixed the call, the service now returns no cards when the call fails or there is no data, and the insight card shows an empty state instead.

What is not finished

  • Prices for SageMaker, Redshift, OpenSearch, Neptune and EMR findings are us-east-1 list prices with no scaling for other regions. We accepted that for now, and each of those findings now says so in its pricing line.
  • The RDS table is built from the MySQL Single-AZ rate, and Aurora instances are priced from the same rows. Aurora's own list rates are higher, and we have not switched to them yet.

What to take from it

If you keep your own cost scripts, search them for defaults:

  1. What does the code do when a price lookup returns nothing? If the answer is "use a number", that number will show up in a report one day.
  2. Do two tables hold the same price? If they can disagree, they will.
  3. If two checks can flag the same resource, does your total add them?
  4. Is a risk labelled as a saving anywhere in your output?

Run it yourself

The CLI is free and open source under the Apache License 2.0. It runs on your machine, reads your AWS account with your own credentials, and sends nothing to CloudWise unless you choose to share a summary. It only reads.

pipx install cloudcostwise          # or: uvx cloudcostwise scan
cloudcostwise scan --profile my-profile --regions all

Read-only and free, in a few minutes. AWS bills the scan's Cost Explorer calls; add --no-cost-explorer to skip them. Start at cloudcostwise.io. The full list of checks is on the waste checks page.

Find this waste in your own account

Run the free scanner on your own machine. Read-only, no signup, and nothing is sent to CloudWise unless you choose to share a summary at the end.

$ pipx install cloudcostwise

What it checksPrefer a dashboard? Start a free hosted scan →