Blob store — AWS
Service mapping
- S3 — the object store. Buckets + keys, strong read-after-write, versioning, multipart upload, presigned URLs, lifecycle rules, event notifications. The default.
- Storage classes (the tiering knob) — Standard (hot), Standard-IA / One Zone-IA (infrequent access, cheaper storage + retrieval fee), Glacier Instant/Flexible Retrieval and Glacier Deep Archive (archive; minutes-to-hours retrieval, cheapest at rest), Intelligent-Tiering (auto-moves objects between tiers by access pattern).
- Durability/replication — S3 handles intra-region durability internally; Cross-Region Replication (CRR) / Same-Region Replication for geo/compliance copies (asynchronous).
- CloudFront — CDN in front of S3 origin (→
content-delivery).
When to pick which
Standard for actively served objects; IA for backups read a few times a month; Glacier/Deep Archive for compliance retention off the request path; Intelligent-Tiering when access is unpredictable and you don't want to manage lifecycle transitions yourself.
Limits / things that bite (verify against current docs)
- Max single object size and the single-PUT vs multipart threshold — large objects must use multipart; multipart part-size minimum (~5 MB) and a max part count cap effective object size.
- Glacier/Deep Archive retrievals take minutes to hours and charge retrieval + request fees — never on a latency path.
- IA/One Zone-IA have a minimum billable object size and minimum storage duration; tiering many tiny objects can cost more than Standard.
- Egress to the internet is billed; CloudFront in front cuts both egress and origin load.
- Request rate scales per prefix — historically high-cardinality random prefixes spread load; monotonic/date prefixes can hotspot.
Provider-specific trade-offs
- Deep S3 integration (IAM, KMS encryption, event → Lambda, S3 Select) is convenient but is lock-in; the S3 API itself is the de-facto standard others emulate.
- Intelligent-Tiering removes lifecycle work but adds a per-object monitoring charge — wasteful for very small objects.
Pitfalls
- Putting Glacier on a read path and timing out requests.
- No lifecycle rule to abort incomplete multipart uploads → orphaned parts billed forever.
- Versioning on with no expiry of non-current versions → silent storage growth.
- Serving popular objects straight from S3 without CloudFront → egress blowup.
- Over-broad/long-lived presigned URLs and bucket policies — scope tight, expire short.