Data Storage — AWS
Service mapping → options
- RDS (Postgres/MySQL/MariaDB) — managed relational; automated backups, read replicas, Multi-AZ failover. The default managed SQL.
- Aurora — RDS-compatible (Postgres/MySQL) with a distributed storage layer: up to 15 low-lag read replicas, fast failover, storage auto-grows. Pick over plain RDS when you need more read scale and quicker failover.
- DynamoDB — managed key-value + document; partition key (+ optional sort key), single-digit-ms reads, on-demand or provisioned capacity, auto-sharded. Maps to key-value / wide-column options.
- ElastiCache — managed Redis/Memcached (see
caching), not a system of record. - Keyspaces (managed Cassandra), Neptune (graph), DocumentDB (Mongo-compatible) — use when you specifically need those models.
Limits / things that bite (verify against current docs)
- DynamoDB: item size capped (~400 KB); a single partition has a throughput ceiling, so a hot partition key throttles even when total capacity is spare — design the PK for even spread. Strongly-consistent reads cost more RCUs and only work on the base table, not global secondary indexes (GSIs are eventually consistent).
- RDS/Aurora: Multi-AZ failover takes seconds to ~a minute; an async read replica can lag and lose the tail if promoted. Connection limits scale with instance size — front with RDS Proxy to survive Lambda/spiky fan-out.
- Cross-AZ data transfer is billable.
Provider-specific trade-offs
- DynamoDB is serverless and scales effortlessly but locks you into its data model and query style (no joins, access patterns must be designed up front); migrating off is real work.
- Aurora decouples compute from storage (cheap replicas) but is AWS-only.
- On-demand DynamoDB is convenient but pricier than well-sized provisioned capacity at steady high load.
Pitfalls
- Choosing DynamoDB then fighting it with scan-heavy/ad-hoc queries it isn't built for — that's a relational workload.
- Hot partition key throttling while the table looks under-provisioned.
- Skipping RDS Proxy and exhausting connections under serverless fan-out.
- Assuming GSIs are strongly consistent.