A vendor once spent forty minutes telling a steering committee I sat on that their platform would give us “a unified semantic layer across the fabric.” Nine people nodded. I nodded. Afterwards, one of the data engineers asked me what happens to the downstream consumers when an upstream team renames a column, and I realised that was the only question in the room worth asking, and that nobody had asked it, including me.

That is the specific failure this reading list is aimed at. Not ignorance — most EAs I know are perfectly competent — but the particular kind of silence that comes from not having the vocabulary to interrogate a claim in real time. Data architecture is now where the interesting decisions in most enterprises happen, partly because every AI-readiness conversation collapses into a data conversation within two meetings. If you cannot push back in that room, you are decorative.

There are three books that fix this. I am not going to pretend there are ten.

First, the vocabulary problem

The reason data architecture reading is hard to recommend is that the field churns its nouns faster than it accumulates ideas. In the decade I have been adjacent to it we have had the data warehouse, the data lake, the lakehouse, the data mesh, the data fabric, and whatever the current deck calls it. Each wave arrived with a book, a conference track, and a consulting practice.

I am not saying these terms are empty. Data Mesh (Zhamak Dehghani) in particular contains a genuinely useful argument about domain ownership of data products, and if your organisation has already committed to the word, read it so you can use the word precisely. But I would not put it in anyone’s first three books, because the durable part of that argument — that data has owners, contracts and consumers, and that centralising it in one team creates a bottleneck — is available from cheaper sources and survives the term going out of fashion. The vocabulary will churn again. Buy the books that survive the churn.

1. Designing Data-Intensive Applications — the non-negotiable

Designing Data-Intensive Applications

· Full review →

There is no substitute and no shortcut. Kleppmann’s book is the only one that gives you the actual mechanics underneath every data-platform claim you will be asked to evaluate, and it is written well enough that you do not need to be an engineer to follow it.

The mistake EAs make is reading it like an engineer: cover to cover, front to back, grinding through storage engines until they stall somewhere in chapter six and never come back. Engineers should read it that way. You should not. You are not going to implement a B-tree. You are going to be asked whether a proposed integration is safe, and what it costs.

So read it selectively, and in this order:

  • Chapter 4, “Encoding and Evolution”, first. This is the chapter that earns the price of the book for an enterprise architect, and almost nobody starts here. Backward and forward compatibility, schema evolution, what Avro and Protobuf actually give you that JSON does not — this is the data contract conversation, stripped of the marketing. Every “we’ll agree an interface between the domains” slide is making promises that this chapter tells you how to keep.
  • Chapter 2, data models. Relational, document, graph, and — critically — why the choice is about the shape of your relationships rather than the fashion of the decade.
  • Chapters 7 and 9, transactions and consistency. Not to become an expert, but to stop conflating eventual and causal consistency, and to know exactly what a team is giving up when they say “we’ll just make it eventually consistent.” That phrase is almost always a deferral of work rather than a design decision.
  • Chapters 11 and 12, streams and the unbundling of databases. The derived-data material has aged into the most important part of the book. Chapter 12’s argument — that a data platform is a composition of specialised stores joined by event logs, not one big database — is now simply how competent platforms are built.

Skim chapters 3, 5, 6 and 8. Know the headlines, skip the internals. You can always go back when a specific decision demands it, and you will.

The thing you actually gain is a trade-off vocabulary. After this book you can say “that’s at-least-once delivery with idempotent consumers, not exactly-once” in a room and watch a vendor’s pitch reorganise itself around the truth. That single move has saved me more architectural pain than any framework I have been certified in.

2. Building Evolutionary Architectures — contracts you can enforce

Building Evolutionary Architectures, 2nd ed

· Full review →

Once you know what a data contract is, the next question is what stops it being broken on a Tuesday afternoon by a team that has never heard of you. The standard enterprise answer is a review gate, and the standard enterprise outcome is that the gate is either bypassed or ignored. I have written at length about why that fails outright once agents are producing change, and data pipelines are where it fails first, because the breakage is silent. Nothing goes red. A dashboard is just quietly wrong for six weeks.

Ford, Parsons, Kua and Sadalage give you the alternative: the fitness function. Express the architectural property you care about as an automated, continuously evaluated test, and it holds regardless of who or what makes the change. Applied to the data layer this is unusually concrete — schema compatibility checks in CI, a test that fails when a producer emits a field the registry does not know about, a freshness budget that breaks the build rather than the quarterly report, assertions on lineage and cardinality. Sadalage’s presence matters here; the database-refactoring lineage shows through, and the 2nd edition is noticeably more useful on data than the first.

Read it second, not first, because the fitness functions worth writing are the ones DDIA teaches you to care about. Reading it the other way round produces tests of things that do not matter.

3. Enterprise Architecture As Strategy — the layer data programmes skip

Enterprise Architecture As Strategy

· Full review →

This is the one nobody expects on a data list, and it is the one I have needed most often.

Almost every failed data programme I have watched failed above the technology line. Someone decided to standardise customer data across business units that have no reason to share customers, or federated ownership to domains in a company that has never in its history let a business unit own anything. The platform was fine. The programme was answering a question the organisation had not agreed on.

The operating-model quadrant in Ross, Weill and Robertson — the one that sorts organisations by how much process standardisation and how much data integration they genuinely require — is the fastest diagnostic I know for that failure. It takes about twenty minutes with a leadership team and it produces an argument, which is the point. A Diversification company running a Unification-shaped master-data programme will spend three years and a great deal of money discovering it is a Diversification company. The case studies are period pieces now, as I have said elsewhere. The diagnostic is not.

Read it third if you are technically inclined, first if you are already senior and the argument you keep losing is about ownership rather than mechanics.

How I would sequence this

  • Pulled into your first data-platform review next month: DDIA chapter 4, then chapters 2, 11 and 12. Two weekends. It is enough to ask the column-rename question.
  • Responsible for governing a platform that already exists: DDIA selectively, then Building Evolutionary Architectures, and write three fitness functions before you read anything else.
  • Running or advising a multi-year data programme: Enterprise Architecture As Strategy first, because the operating-model question decides whether the rest of the work is even coherent.

Bottom line

Three books, read in the order the argument needs rather than the order they were published: Kleppmann for the mechanics, Ford et al for the enforcement, Ross et al for the organisational question underneath both. Everything else in this category — and there is a great deal of it — is either a restatement of one of these three or a vocabulary that will be replaced before your programme finishes.

If you read only one, read DDIA, selectively, starting at chapter 4. It is the difference between nodding along in the fabric meeting and being the person who asks what happens when someone renames a column.

Data architecture reading — quick answers

What is the best data architecture book for an enterprise architect?
Designing Data-Intensive Applications by Martin Kleppmann. Read it selectively rather than cover to cover: chapter 4 on encoding and evolution first, then data models, then the streaming and derived-data chapters. That gives you the trade-off vocabulary to interrogate a data-platform proposal without being the engineer who builds it.
Do I need to read Designing Data-Intensive Applications cover to cover?
No, and most enterprise architects who try it that way stall around the storage-engine chapters. Skim chapters 3, 5, 6 and 8 for the headlines and read 2, 4, 7, 9, 11 and 12 properly. Go back to the internals when a specific decision demands it.
Should I read a data mesh or data fabric book?
Only if your organisation has already committed to the term and you need to use it precisely. The durable idea — that data has owners, contracts and consumers, and centralising it in one team creates a bottleneck — is available from Kleppmann and Ford et al, and survives the vocabulary going out of fashion.
What should I read about data contracts specifically?
Chapter 4 of Designing Data-Intensive Applications for what a contract actually guarantees, then Building Evolutionary Architectures for how to enforce it. A data contract that is not expressed as an automated check in a pipeline is a document, not a contract.
Why is Enterprise Architecture As Strategy on a data reading list?
Because most data programmes fail above the technology line. The operating-model quadrant tells you how much process standardisation and data integration your organisation genuinely needs, which decides whether a master-data or federated-ownership programme is coherent before anyone chooses a platform.