Amazon S3 (Simple Storage Service) is AWS object storage. You put files (objects) in buckets. Each object has a key (path-like name), data bytes, and metadata. Data engineers use S3 as the landing zone and data lake: raw dumps, Parquet tables, logs, backups.
Problem first. A warehouse of CSV exports and Spark job outputs needs cheap, durable storage that many tools can read. Attaching giant disks to one server does not scale. S3 is shared object storage with 11 nines of durability design.
Producer / ETL job
|
v
s3://lake-prod/
raw/orders/dt=2026-09-05/part-000.parquet
curated/fct_orders/...
logs/spark/...Core ideas
- Bucket = top-level container (globally unique name)
- Object = file + key + metadata
- No real folders; prefixes like
raw/orders/are a naming convention - Access via IAM policies, bucket policies, and (optionally) ACLs
Storage classes (interview table)
| Class | When to use | Access pattern | Cost shape | |---|---|---|---| | S3 Standard | Hot data, frequent reads | Milliseconds | Higher storage, low request cost | | S3 Intelligent-Tiering | Unknown / changing access | Auto-moves tiers | Small monitoring fee | | S3 Standard-IA | Infrequent but quick access | Milliseconds + retrieval fee | Cheaper storage | | S3 One Zone-IA | Re-creatable infrequent data | One AZ only | Cheapest IA, less resilient | | S3 Glacier Instant Retrieval | Archive, rare but need ms access | Milliseconds | Low storage, higher retrieval | | S3 Glacier Flexible Retrieval | Cold archive (minutes–hours) | Restore then read | Very low storage | | S3 Glacier Deep Archive | Long-term compliance | Hours to restore | Lowest storage |
Lifecycle example
Keep last 30 days in Standard, move older prefixes to IA, then Glacier after 90 days. Use lifecycle rules so you do not manage this by hand.
Interview tip: Say "S3 = durable object store for lakes," then name 2–3 classes and when you would lifecycle data down. Mention prefixes and IAM, not "folders."