Cloud Databases — Quick Reference
Provider landscape
| Provider | Relational | NoSQL | Known for |
|---|
|---|---|---|---|
| AWS | RDS, Aurora | DynamoDB | Broadest ecosystem, Aurora storage-compute separation |
|---|
| Azure | Azure SQL, Azure DB for PostgreSQL | Cosmos DB | Cosmos DB multi-model, multi-region writes |
| GCP | Cloud SQL, AlloyDB | Firestore, Bigtable | Spanner — globally-consistent relational scale |
| Vendor-neutral | Self-managed Postgres/MySQL | MongoDB Atlas, Redis Cloud | Portability across clouds |
Multi-AZ vs. Read Replica
|---|---|---|
| Purpose | High availability / failover | Read scaling |
|---|
| Replication | Synchronous | Usually asynchronous |
| Queryable directly? | No | Yes |
| Failover | Automatic | Manual promotion |
Relational vs. NoSQL — decision shortcut
Need ACID + joins + well-understood schema → managed relational
Need massive horizontal scale + ms latency → managed NoSQL
Need global multi-region WRITES specifically → Spanner / Cosmos DB
(not just multi-region reads — a genuinely distinct, higher-cost category)
DynamoDB partition key design
Low cardinality (e.g. a status field, 3 values) → hot partition risk
High cardinality (e.g. user_id) → even distribution
Time-series pattern to avoid hot partitions:
partition_key = f"{device_id}#{date.today().isoformat()}"
AWS RDS — quick commands
bash
# Provision with Multi-AZ and automated backups
aws rds create-db-instance --db-instance-identifier myapp-prod \
--engine postgres --multi-az --backup-retention-period 7
# Check replication lag on a read replica
aws cloudwatch get-metric-statistics --namespace AWS/RDS \
--metric-name ReplicaLag --dimensions Name=DBInstanceIdentifier,Value=myapp-replica
# Manually promote a read replica (for DR failover)
aws rds promote-read-replica --db-instance-identifier myapp-replica
Cost model — pricing dimensions
Compute — instance size / capacity units
Storage — per GB-month
I/O/requests — per request (NoSQL) or per I/O operation
Data transfer — cross-AZ/region egress
DynamoDB specifically:
On-demand — pay per request, no capacity planning, no throttling risk
Provisioned — cheaper at steady predictable load, throttles if under-sized
DR design — RTO/RPO reference
| Architecture | Typical RTO | Typical RPO | Relative cost |
|---|
|---|---|---|---|
| Cross-region read replica + manual promotion | Minutes–tens of minutes | Seconds–minutes | Lower |
|---|
| Multi-region write (Cosmos DB / Spanner) | Near-zero | Near-zero | Significantly higher |
Migration checklist
[ ] Schema/compatibility assessment run BEFORE committing to a timeline
(AWS SCT or equivalent — proprietary features are the real risk,
not the bulk data transfer)
[ ] Migration tool validation/comparison report run before cutover
[ ] Cutover strategy chosen deliberately: big-bang (simpler, has a
downtime window) vs. dual-write (no downtime, more complex)
[ ] Rollback plan defined before cutover, not improvised during it
