yanacuti1121 OBEY Designing Data-Intensive Applications by Martin Kleppmann
When to use
Use when data correctness, durability, or distributed write semantics matter more than local code style.
Primary bias to correct
Hidden data contracts are still contracts.
Decision rules
- State the source of truth, consistency expectation, durability point, visibility point, retry semantics, and evolution path for every important data change.
- Choose data models, storage, indexes, replication, partitioning, transactions, queues, streams, and APIs from workload, access pattern, consistency, reliability, maintainability, and operational cost.
- Treat caches, indexes, projections, search copies, denormalized data, and materialized views as derived data with staleness, lag visibility, repair, and rebuild paths.
- Make retried, replayed, queued, batch, stream, and event-driven work idempotent or transactional; reject casual exactly-once claims.
- Treat schemas, encodings, service APIs, messages, logs, and events as versioned contracts that must survive old code, old data, rolling upgrades, and in-flight messages.
- Assume distributed uncertainty: crashes, partial writes, timeouts, duplicate messages, reordered events, stale replicas, lag, clock error, pauses, stale leaders, and unknown success.
- Match replication, partitioning, isolation, transactions, and coordination to the invariant; do not rely on follower freshness, quorum formulas, weak isolation, wall-clock order, or ad hoc leadership without proof.
Trigger rules
- When adding retries, jobs, consumers, queues, CDC, event sourcing, or stream processing, prove duplicate, replay, ordering, side-effect, and recovery safety.
- When changing schemas, APIs, messages, events, enum values, or status meanings, plan backward and forward compatibility plus migration, bootstrap, or rebuild paths.
- When reading from replicas or partitioning data, define staleness, routing, hot-key, ordering, rebalancing, and cross-partition behavior.
- When using locks, leases, timestamps, leadership, majorities, or coordination services, define the fault model, quorum/session semantics, stale-authority behavior, and fencing.
Final checklist
- Clear owner and source of truth?
- Explicit consistency, durability, visibility, and staleness semantics?
- Safe under retry, replay, duplicate delivery, reordering, and unknown success?
- Compatible across old data, old code, new code, and messages in flight?
- Isolation, replication, partitioning, transactions, and coordination checked against actual invariants?