@dzhechkov/skills-book-ddia

Decision-moment skills for building data-intensive systems.

Inspired by Designing Data-Intensive Applications by Martin Kleppmann (O'Reilly) — the canonical reference on data systems. This package is an ORIGINAL, rephrased distillation of the field's decision methodology; it reproduces none of the book's text. It teaches WHEN to reach for which data-systems technique (replication topology, partitioning, transaction isolation, consensus, encoding evolution) with original criteria and tradeoff tables. For the full treatment, the ideas' depth, and the author's own words, read the book — this is a working companion, not a substitute. trust_tier: 0 (machine-distilled): verify a decision against primary sources before relying on it.

What it does

Makes an AI coder apply the book's data-systems methodologies at the real design decisions — not summarize the book. Each skill activates on a decision moment and gives concrete criteria, tradeoff tables, key facts/formulas with page anchors, anti-patterns, and cross-links.

The 10 decision-moment skills (116 verified Knowledge Units)

| Skill | Decision | Chapters | |-------|----------|----------| | ddia-reliability-scalability-foundations | reliability / scalability / maintainability framing | 1 | | ddia-data-model-selection | relational vs document vs graph | 2 | | ddia-storage-engine-tradeoffs | LSM vs B-tree, OLTP vs OLAP | 3 | | ddia-encoding-and-schema-evolution | serialization + backward/forward compat | 4 | | ddia-replication-topology-choice | single / multi / leaderless + consistency guarantees | 5 | | ddia-partitioning-strategy | hash/range partitioning, rebalancing, secondary indexes | 6 | | ddia-transaction-isolation-choice | isolation levels, SSI, when you need a transaction | 7 | | ddia-distributed-consistency-consensus | distributed faults, linearizability, consensus | 8–9 | | ddia-batch-and-stream-processing | batch vs stream, dataflow, exactly-once | 10–11 | | ddia-deriving-data-and-integration | derived data, system-of-record, end-to-end integration | 12 |

Install (owner-local)

dz init --target claude-code --select ddia-replication-topology-choice   # one decision
# then: «проектирую репликацию — single или multi-leader?» → the skill activates

Each skill ships its source Knowledge Units in references/knowledge-units.md (in-pack deep-lookup). The full KB is queryable: dz recall --books --book vysokonagruzhennye-prilozheniya "<query>".

Usage scenarios

Install the pack once (owner-local — it's private, so from the monorepo: dz init --target claude-code --select <the 10 ddia- ids>), then just describe your task to Claude Code in plain language — the agent auto-activates the right skill(s). No skill ids to memorize. Here are the situations where this pack pays off, with example prompts you can copy and adapt:

1. Design a new data-intensive system from scratch

Situation: greenfield — you must pick a data model, storage engine, replication, and partitioning before writing much code, and want the tradeoffs made explicitly.

«Используй скиллы ddia: проектирую систему аналитики событий на 50k RPS — какую модель данных, движок хранения, репликацию и партиционирование выбрать?»

What happens: the agent walks the coupled decisions in order (model → storage engine → replication → partitioning), applying DDIA's criteria at each and flagging where one choice constrains the next (e.g. leaderless replication → quorum math in consistency).

2. Make one architecture decision, right now

Situation: you hit a specific fork and want the principled, cited answer.

«Проектирую репликацию — single-leader, multi-leader или leaderless?» (EN: "single-leader, multi-leader, or leaderless replication?")

What happens: ddia-replication-topology-choice activates, weighs simplicity vs write-availability vs conflict-resolution cost against your context, and links the coupled decision (ddia-transaction-isolation-choice / consistency) — with page-anchored citations.

3. Review / red-team a data architecture

Situation: an existing design review — you want risky choices surfaced before they ship.

«Используй скиллы ddia: отревьюь эту схему — multi-leader между дата-центрами + read-committed — где риски?»

What happens: the agent flags the failure modes the book warns about (multi-leader write conflicts without a resolution strategy, isolation anomalies under contention, hot partitions) with the rationale.

4. Choose consistency & isolation for a requirement

Situation: a correctness requirement ("no double-spend", "read-your-writes") and you need the right isolation/consistency level, not the strongest-by-default.

«Нужна ли мне линеаризуемость здесь, или достаточно причинной согласованности? И какой уровень изоляции транзакций?»

What happens: ddia-distributed-consistency-consensus + ddia-transaction-isolation-choice map the requirement to the minimal sufficient guarantee and its cost, distinguishing what actually needs consensus from what doesn't.

5. Evaluate a database / storage migration

Situation: picking or migrating a datastore and weighing engine tradeoffs + schema evolution.

«Используй скиллы ddia: переходим с Postgres на что-то под тяжёлую запись — LSM vs B-tree, и как не сломать совместимость схемы?»

What happens: ddia-storage-engine-tradeoffs (LSM vs B-tree write/read/space amplification) + ddia-encoding-and-schema-evolution (backward/forward compatibility) frame the migration's real costs.

Provenance

sources.json: upstream_type: book, ISBN, per-skill derived_from KU ids, 100% verified ratio, digitizer corpus_version. No origin block (the book is the immutable upstream).