Memory hook: Access pattern before product name.
Must remember
- Cloud SQL offers managed MySQL, PostgreSQL and SQL Server; AlloyDB targets demanding PostgreSQL-compatible workloads with managed architecture and read pools. Check extension and feature compatibility rather than assuming all PostgreSQL systems are interchangeable.
- Spanner provides horizontally scalable relational transactions with strong consistency. Bigtable suits high-throughput key/range access; Firestore suits document-centric applications; Memorystore supplies low-latency caching. BigQuery is an analytical warehouse.
- Assess transaction shape, joins, consistency, data size, peak concurrency, working-set memory, IOPS, throughput and growth. Benchmark representative traffic rather than sizing solely from database file size.
- Managed services reduce operations but impose supported engines, versions, extensions and configuration boundaries. Self-managed or partner/bare-metal offerings can meet requirements outside those boundaries at higher operational responsibility.
- Include region availability, data residency, organization policy, licensing, network egress and backup retention in the design. Multiple databases may be appropriate when each serves a distinct access pattern.
- Vector retrieval can support AI grounding, but choose index type, filtering, distance metric and embedding lifecycle deliberately. A vector index does not replace transactional correctness or permission checks.
Choose under exam pressure
| Requirement | Choice and reason |
|---|---|
| Global relational transactions with horizontal growth | Evaluate Spanner with schema and latency testing. |
| Existing compatible PostgreSQL application | Compare Cloud SQL and AlloyDB before rewriting for a different data model. |
Traps
- NoSQL is not a synonym for no schema design.
- A cache is not automatically a durable system of record.
Active recall
1. Why measure working-set memory?
To understand cache effectiveness and the I/O required under real traffic.
2. Which service fits key-range scans at huge throughput?
Bigtable, with a row-key design that avoids hotspots.
3. Why assess extensions before migration?
A managed target may not support the source extension or behavior.
4. What makes BigQuery a poor default for OLTP?
Its architecture and pricing target analytical workloads rather than many small transactional requests.
5. What must accompany vector search?
Correct embeddings, evaluation, data freshness and authorization filters.