Further Reading: Cloud Cost Optimization

Sources are tagged Tier 1 (confident it exists, recommended without reservation) or Tier 2 (real and worth seeking, but confirm the current edition, version, or URL yourself).

Two warnings. First, the highest-value reading in this chapter is your own bill, and almost nobody has read theirs line by line. Everything below is a distant second. Second, this literature has an unusually high vendor content, and vendor cost material has a structural bias: it explains how to spend their product efficiently and rarely how to spend less on it.

The one that is genuinely required

Your cloud provider's pricing pages, for the three or four services that dominate your bill. Not a summary, not a blog post — the pricing page, including the footnotes. You need to know the unit, the tiering, and what is billed separately (requests, egress, cross-AZ transfer, early deletion). §33.2's five-minute check is this, and it is the highest return-on-time item in the chapter. Tier 1.
Your own bill, at line-item granularity, exported to somewhere you can query it. AWS Cost and Usage Report, GCP billing export to BigQuery, Azure Cost Management export. The console's summary view is not enough — Case Study 1's attribution needed the line items. Tier 1.
Your warehouse's usage views. Snowflake ACCOUNT_USAGE.WAREHOUSE_METERING_HISTORY and QUERY_HISTORY; BigQuery INFORMATION_SCHEMA.JOBS with total_bytes_billed. Cost per query, per user, per pipeline, already there and almost never read. Tier 1.

On the discipline

J.R. Storment and Mike Fuller, Cloud FinOps (O'Reilly, 2nd ed. 2023). The standard text. The chapters on allocation, showback versus chargeback, and unit economics are the relevant ones — §33.5 and §33.11 are compressions of them. Skim the organizational-maturity material unless you are building a function. Tier 1.
The FinOps Foundation's framework and its "capabilities" list. Free, vendor-neutral, and useful as a checklist of things you may not have considered. Read it the way §30's DAMA-DMBOK entry recommends — to find gaps, not to build a program, since it presents every capability as necessary. Tier 1.
Corey Quinn's Last Week in AWS newsletter and writing. Opinionated, frequently funny, and more honest about cloud pricing perversity than any other regular source. The recurring theme — that data transfer pricing is the least rational part of the bill — is §33.7's cross-region line, argued at length by someone who does this for a living. Tier 2 — a newsletter; find the current archive.

On the engineering that actually saves money

Chapter 9 of this book, on partitioning and file size. Not a deflection: partition pruning is the single highest-value optimization in this book, and Chapter 1's $1,374,269 is one instance of it.
Your engine's documentation on reading a query plan. Spark's EXPLAIN FORMATTED, Snowflake's query profile, BigQuery's execution details, DuckDB's EXPLAIN ANALYZE. The scan node and its byte count is the number that matters, and reading a plan remains the highest return-on-time skill in this book. Tier 1 — versioned per engine.
BigQuery's dry-run documentation and Snowflake's EXPLAIN. §33.6's pre-flight estimate. Ten minutes of reading, and it is the mechanism behind Case Study 2's pull-request bot. Tier 1.
The AWS, GCP, and Azure documentation on Spot / Preemptible / Spot VMs, particularly the interruption-rate data. §33.10. The interruption rates are published per instance type and region and are the input to the only honest spot decision. Tier 1.

On commitments

Your provider's Savings Plan / Committed Use Discount calculator, run twice: once against current usage and once against the floor you are confident of in two years. §33.10's whole argument is the difference between those two numbers, and the tools will happily show you only the first. Tier 1.
Anything careful on the economics of sunk cost and its behavioural effects. §33.10's claim — that a commitment stops teams pursuing savings they cannot benefit from — is an instance of a well-documented effect, and one clear treatment will make it easier to argue for. Tier 2 — the behavioural-economics literature; one chapter suffices.

On monitoring cost

Your provider's budget and anomaly-detection services — AWS Cost Anomaly Detection, GCP budget alerts, Azure Cost Alerts. They are real and they are coarse: they detect an account-level anomaly and cannot tell you which DAG. Case Study 2's per-pipeline check is the thing they do not do. Tier 1.
Chapter 25 and Chapter 26 of this book. Case Study 2's finding is that the entire monitoring stack built there measures time, correctness, and volume, and that none of it measures money. The fix is a Chapter 25-shaped monitor on a new dimension.
OpenCost and Kubecost, if you run on Kubernetes. Allocation for shared clusters is the hardest attribution problem in this chapter and these are the credible open-source answers. Tier 1.

On unit economics

Anything on SaaS unit economics and cost of goods sold. §33.11's cost-per-order and percent-of-GMV are standard business metrics that data teams rarely apply to themselves. The transferable idea is that a total is not a number anyone can act on, and the framing is well-established outside engineering. Tier 2 — a large literature; one good primer.
Public engineering write-ups on cost-per-unit at scale — Dropbox's storage migration, Netflix's and Pinterest's cost-efficiency posts, Slack's warehouse work. Read them for the decomposition, not the conclusions, which are specific to workloads unlike yours. Tier 2 — blog posts move.

Practice

code/cost_model.py in this chapter. Three meters, an attributed bill, seven sized wastes, and a payoff-per-week ranking, with forty-eight self-checks, all deriving from one rate card. Exercise 33.11's first task — replace the rate card with yours and re-run everything — is the most useful hour in the chapter.
Estimate three queries before running them. Exercise 33.4. Most people are wrong by more than 10× on the first and within 2× by the third, and the skill is almost entirely calibration.
Compute your unattributed share. Exercise 33.5. Report it honestly, including if it is embarrassing — the embarrassing number is the valuable one.
Ask, of every compute resource you run: what fraction of its billed time is it working? §33.7. This one question found 83.3% of Kestrel's waste.

A note on what to be skeptical of

Any tool that promises to cut your bill by a percentage. The percentage is achievable and the tool is rarely what achieves it; §33.7's top three items are a settings change, a schedule, and a refresh interval, and no product was involved.

Any cost comparison between vendors that uses list prices. Nobody at scale pays list, discounts vary enormously, and the comparison is usually published by whichever vendor wins it.

Storage-tier recommendations that omit retrieval cost. Glacier is dramatically cheaper to store and can be dramatically more expensive to read, and a recommendation that models only the first half is worse than none.

And any cost figure without a unit. After this chapter, "we spend $30,000 a month" should prompt the same reflex as Chapter 32's "AUC 0.83": per what?