Data is where state lives and who is allowed to reach it. Of the five pillars it is the hardest to change: compute can be replaced and networks reconfigured, but moving or restructuring live data is slow and risky. Data decisions therefore deserve the most deliberation early.
| Type | Holds | Typical use |
|---|---|---|
| Managed database | Structured, queryable state | Transactional data, relational or document models |
| Object storage | Large unstructured blobs | Files, backups, media, data lakes |
| Cache | Hot, ephemeral copies | Reducing read latency and database load |
How state is structured, evolved, and indexed is covered in Data Architecture; caching strategies and invalidation in Caching.
Where data physically lives is a legal and architectural constraint, not only a latency concern. Regulations may require that certain data stay within a jurisdiction. Residency requirements influence region selection, replication topology, and which managed services are usable. Decide them before choosing products.
Cloud data services replicate data for durability and availability. Replication introduces more than one copy of the same data, which raises the question of what a read guarantees (see Consistency Models). Backups and recovery targets are part of Reliability.
| Decision | Benefit | Cost |
|---|---|---|
| Managed database | No patching, backups, or failover to operate | Less control; dependency on the provider |
| Object storage for large blobs | Cheap, durable, scales without planning | Higher latency than block or in-memory storage |
| Proprietary data service | Powerful features, low operational cost | Migration away is expensive (lock-in) |
| In-region replication | Higher durability and availability | Storage cost and write latency increase |
- State in a single zone: data confined to one availability zone shares that zone's failure. Durability requires replication across zones (see Reliability).
- Residency overlooked: discovering a jurisdiction requirement after choosing a region forces a costly migration.
- Lock-in through proprietary formats: data stored in a provider-specific service is the most expensive thing to migrate later. Weigh the feature benefit against exit cost.
- No verified backups: a backup that has never been restored is an assumption, not a recovery plan.