Cloud basics interview questions for data and AI roles

Storage

1. Object storage versus a database — when do you use each?

Object storage (such as Amazon S3, Azure Blob Storage or Google Cloud Storage) holds files — raw exports, images, Parquet datasets, model artefacts — cheaply and durably, addressed by a key. A database serves structured queries and transactions with low latency. A typical data platform lands raw files in object storage and loads or queries them from a warehouse.

Trap: treating object storage like a file system you edit in place. Objects are usually written whole and replaced, not appended line by line.

2. What are storage classes or tiers?

Cheaper tiers trade retrieval speed or retrieval cost for lower storage price. Frequently read data stays in the standard tier; old raw logs kept for compliance can move to an archive tier through a lifecycle rule.

Trap: archiving data that a daily job still reads, which can raise the bill through retrieval charges.

3. Why do data teams prefer Parquet over CSV in object storage?

Parquet is columnar and compressed, so queries read only the columns they need and scan far fewer bytes. It also stores data types, avoiding the "is this a date or a string?" problems of CSV.

Compute

4. Virtual machine, container, or serverless function?

A virtual machine gives full control and suits long-running or specialised workloads. Containers package an application with its dependencies so it runs the same everywhere and can be scheduled by an orchestrator. Serverless functions run short tasks on demand with no server management and bill per invocation, but have time and memory limits.

Trap: choosing a serverless function for a job that runs for hours. Check the platform's execution limit first.

5. What is autoscaling?

Adding or removing instances automatically based on load, such as CPU or queue length. It saves money off-peak and absorbs peaks, but needs sensible minimums and maximums so a bug cannot scale costs without limit.

Identity and access

6. What is IAM, and what is least privilege?

Identity and access management defines who (users, services) can do what on which resources. Least privilege means granting only the permissions a task needs — a nightly job that reads one bucket gets read access to that bucket, not administrator rights.

Trap: putting long-lived access keys in code or notebooks. Use roles or managed identities, and a secrets manager for anything that must be a secret.

7. A colleague made a storage bucket public "just for a quick share". What do you do?

Make it private again, share through a time-limited signed URL if needed, check the access logs for what was downloaded while it was open, and tell whoever owns security. Public buckets are a common cause of data leaks.

Networking

8. What are a VPC and a private subnet?

A virtual private cloud is your isolated network in the provider. Resources in a private subnet have no direct route from the internet; databases usually live there, reachable only from application servers or through a controlled gateway.

9. What is data egress, and why does it matter?

Data transferred out of the provider's network, or sometimes between regions, is usually charged, while inbound transfer is often free. A design that copies large datasets out every day can cost more in transfer than in storage.

Regions and reliability

10. How do you choose a region?

Close to users for latency; where data residency obligations and customer contracts require; where the services you need are available; and at acceptable price. For Indian personal data, check the organisation's legal guidance rather than assuming.

11. Availability zones versus regions?

Zones are separate data-centre locations within one region; spreading across zones protects against a single facility failing. Multi-region designs protect against a region-wide outage but cost more and are more complex.

Cost

12. Estimate the monthly cost of a small data pipeline.

Say you would list the cost drivers — storage, compute hours, queries scanned, egress — and multiply by placeholder unit prices from the provider's calculator:

# Placeholder unit prices in rupees; replace with the provider's current pricing page
price = {"storage_gb_month": 2.0, "compute_hour": 8.0, "query_tb_scanned": 450.0, "egress_gb": 7.0}
usage = {"storage_gb_month": 500, "compute_hour": 2 * 30, "query_tb_scanned": 3, "egress_gb": 200}

lines = {k: usage[k] * price[k] for k in price}
total = sum(lines.values())
for k, v in sorted(lines.items(), key=lambda kv: -kv[1]):
    print(f"{k:<18} Rs {v:>8,.0f}  ({v / total:.0%})")
print(f"{'total':<18} Rs {total:>8,.0f}")

With these placeholders, egress and query scanning together are about two-thirds of the bill, and storage just under a quarter — the usual lesson that how data is read and moved matters more than how much is stored. Converting CSV to partitioned Parquet reduces the scan line directly.

Trap: quoting remembered prices. Interviewers prefer "I'd check the pricing calculator" plus a correct structure.

Preparing

Create a free-tier account with one provider, upload a CSV to object storage, grant a role read-only access, query it with the provider's serverless SQL tool, and delete everything afterwards. One hands-on hour answers most of these questions better than reading.

Where pipelines run is half the story; how they rerun safely is in the data engineering ETL bank. Cloud deployment is part of the production-engineering phase of FDE for Freshers, described on the forward deployed engineer hub.