Choosing Cloud Data Storage
“Where should this data go?” is not one question. A video object, a customer order, an offline mobile document, and a high-volume sensor stream have different access patterns and correctness requirements.
Start with the shape of the workload, then select a service.
Identify the Data Contract
Ask five questions:
- Is the value an opaque object or structured records with fields and relationships?
- Which queries must be fast?
- Do several writes need to succeed or fail as one transaction?
- How much data and request throughput must the system sustain?
- How often will old data be read or changed?
Common Storage Families
| Family | Good fit | Important constraint |
|---|---|---|
| Object storage | Images, video, backups, archives, and processing artifacts | Applications normally replace whole objects rather than update rows inside them |
| Managed relational database | Structured records, SQL queries, joins, and transactions | A single-instance architecture eventually has scaling limits |
| Distributed relational database | SQL and strong transactions at larger geographic or throughput scale | More cost and operational constraints than a conventional database |
| Document database | Application-shaped documents, mobile synchronization, and flexible schemas | Cross-document relationships are less natural than in a relational model |
| Wide-column database | Very large, high-throughput or time-ordered datasets | Query patterns should be designed around keys; relational joins are not the model |
One application may use several families. For example, a relational database can store an upload's ownership and status while object storage holds the binary content.
Access Frequency Is a Cost Signal
Object stores often provide classes for frequently accessed, infrequently accessed, and archival data. Colder classes can reduce storage cost while increasing retrieval charges or minimum retention commitments.
Classify data from its expected lifecycle, not merely its age. Lifecycle rules can move or delete objects automatically, but an incorrect rule can make a recovery slow or remove data that still matters.
Placement and Failure Behavior
Storage location affects latency, residency, and recovery from infrastructure failures. Replication across zones or regions can improve durability, but the application must understand the service's consistency and failover behavior. See cloud regions and zones for the placement trade-off.