The building blocks and when to reach for each
A map of storage, caching, messaging and compute choices, with the question that picks between them.
You do not need to know every product. You need to know the categories, what each is good and bad at, and the managed equivalent on the cloud you know. Use the cloud you are fluent in (for you, Azure) and translate when asked.
| Need | Reach for | Watch out for | Azure / AWS examples |
|---|---|---|---|
| Transactions, joins, constraints | Relational database | Write scaling beyond one primary; schema migrations at size | Azure SQL, PostgreSQL Flexible Server / RDS, Aurora |
| Huge key-value or wide-column scale, predictable access | Partitioned NoSQL | Hot partitions; you must design for the query | Cosmos DB / DynamoDB |
| Blobs: files, model artefacts, logs | Object storage | Listing is slow; no in-place edits | Blob Storage / S3 |
| Sub-millisecond reads of hot data | In-memory cache | Invalidation, stampedes, memory cost | Azure Cache for Redis / ElastiCache |
| Decoupling producers and consumers | Queue | Duplicates (at-least-once), poison messages | Service Bus, Storage Queues / SQS |
| Ordered, replayable event history | Log / stream | Partition count fixes parallelism; retention cost | Event Hubs / Kinesis, Kafka |
| Full-text and filtered search | Search index | It is a second copy; keep it in sync | Azure AI Search / OpenSearch |
| Semantic similarity | Vector index (often inside the DB or search service) | Recall versus latency knobs; re-embedding on model change | Azure AI Search vectors, pgvector / OpenSearch, managed vector DBs |
| Long-running, multi-step work | Workflow engine / durable functions | Determinism rules; versioning in-flight workflows | Durable Functions, Logic Apps / Step Functions, Temporal |
Three questions that pick storage
- What are the top three queries, and does each have a key you can partition on?
- What must be atomic together? Keep that inside one partition or one transaction boundary.
- How fast does the data grow, and how long must it live? That decides tiering and archival.
Your Kubernetes and Azure depth is a genuine advantage in the operations part of a design round. Use it in deep dives (rollout strategy, autoscaling signals, pod disruption budgets, zone spread), not as a list of product names.
Practice
Written for this site to drill the lesson. Not reported interview questions.
- For a ride-hailing app, name the storage for trips, driver locations, payments and receipts, with one sentence each on why.
- Explain to a non-engineer why a queue lets two teams ship independently.
- Pick one row of the table you have run in production. Describe the failure you hit and how you would design around it now.
Check yourself
What is the main risk of a partitioned NoSQL store?
Hot partitions, and having to design the data model around known queries.
Why is a search or vector index a consistency problem?
It is a second copy of the data that must be kept in sync with the source of truth.
Saved in this browser.