Blob store — Azure
Service mapping
- Azure Blob Storage — the object store, inside a storage account → containers → blobs. Block blobs are the object-storage workhorse (uploaded as blocks then committed — Azure's multipart); append blobs suit log appends; page blobs back disks (not general object storage).
- Access tiers (the tiering knob) — Hot, Cool, Cold, and Archive (offline; rehydrate to an online tier over hours before reading). Set per-blob or as an account default.
- Redundancy (the durability knob, set on the account) — LRS (replicas in one datacenter), ZRS (across availability zones), GRS/GZRS (asynchronous copy to a second region), RA-GRS (read access to the secondary). Pick by failure domain you must survive.
- Azure CDN / Front Door — CDN in front of the blob origin (→
content-delivery). - SAS (Shared Access Signature) — Azure's signed-URL equivalent: scoped, time-limited, per-blob or per-container grants.
When to pick which
Hot for actively served blobs; Cool/Cold for backups and infrequently read data (cheaper storage, higher access cost, minimum retention periods); Archive for compliance retention off the request path. LRS for cheapest single-region; ZRS for AZ resilience; GRS/GZRS for region loss.
Limits / things that bite (verify against current docs)
- Block-blob max size is a function of max block size × max block count; large blobs require the block (multipart) path.
- Archive is offline — you must rehydrate (hours) to Hot/Cool before a read; plan it as async.
- Cool/Cold/Archive have minimum storage-duration charges and per-GB retrieval costs; churny or tiny blobs can cost more in a cool tier.
- Redundancy is largely an account-level setting — mixing requirements may mean multiple accounts.
- GRS replication to the secondary region is asynchronous → possible data loss window on failover.
Provider-specific trade-offs
- Tight integration with Entra ID, Event Grid (blob events), and lifecycle management policies; convenient but Azure-specific.
- Account-scoped redundancy/throughput limits make the storage-account boundary a real design unit, unlike S3's flatter bucket model.
Pitfalls
- Reading from Archive without budgeting rehydration latency → timeouts.
- No lifecycle policy to delete old versions/snapshots or tier down cold data → silent growth.
- Over-broad or long-lived SAS tokens (especially account-level SAS) — scope to blob + short expiry.
- Serving hot blobs from origin without Front Door/CDN → egress and throughput pressure.