Distributed Systems Thinking: Architectural Judgment for the AI-Assisted Engineer - Softcover

Inhaltsangabe

There are no wrong answers in architecture, only expensive ones.

Ask a model for a distributed architecture and you have one in seconds. It will be plausible. It will also be quiet about the five things that decide whether the system survives production: what happens when a dependency fails, how tenants stay apart, what breaks when the schema changes, what a single answer costs, and how stale data gets invalidated.

Generated designs are confident by construction. Pricing them is not something the model does for you.

DISTRIBUTED SYSTEMS THINKING is a working course in the judgment that is still yours. Fourteen chapters, each built around a decision you will actually face, worked end to end on one real system instead of illustrated with toys.

That system is DocuMind: a document-intelligence product with five tenants, about 900 questions a day, and an entirely ordinary set of problems. Over fourteen chapters it survives a retry storm that runs for three hours and forty minutes, a poisoned document that reaches the corpus, a model upgrade that quietly moves answer quality, and a schema decomposition that almost cannot be reversed. You watch every decision get made, priced, and written down.

What you will be able to do

• Price a tradeoff rather than argue about it, using a method that fits on one page
• Find the real boundary in a system, which is rarely the one drawn on the diagram
• Choose a consistency level and state the business consequence of the gap you just bought
• Compose resilience properly: timeouts, retries with jitter, circuit breakers, bulkheads, and honest degradation instead of silence
• Instrument a non-deterministic system with golden sets, graders, and drift alarms
• Review an AI-generated design against the five clusters where such designs reliably go quiet
• Write a decision record that still makes sense to whoever inherits it

Three chapters take the AI stack seriously on its own terms. One separates what is genuinely new about it, which is cost as a first-class design force, non-determinism as a property you pay to reduce, and autonomy measured by blast radius, from the large remainder that is classical distributed systems wearing a new noun. Another turns the review of generated designs into a repeatable workflow. The last carries the whole method out to edge deployments, 5G and 6G networks, and orbit, where every instinct transplanted without repricing fails.

Forty-five figures. Every reference verified against a live source.

Who this is for

Senior engineers and architects shipping LLM features who suspect the design review is missing something. Engineers who can build anything and are not yet sure how to decide. Tech leads who have to defend a choice to people holding a budget. Students who want the reasoning that the textbooks state as finished fact.

Avishek Nag is an Associate Professor at University College Dublin, where he teaches and researches distributed and networked systems.

The distinction is intellectual engagement, not tool use. When answers are free, judgment is the scarce skill.

Die Inhaltsangabe kann sich auf eine andere Ausgabe dieses Titels beziehen.