Blob store — GCP
Service mapping
- Cloud Storage (GCS) — the object store. Buckets + objects, strong global read-after-write, object versioning, resumable + multipart uploads, signed URLs, lifecycle (Object Lifecycle Management). The default.
- Storage classes (the tiering knob, set per-bucket default or per-object) — Standard (hot), Nearline (≈ once-a-month access), Coldline (≈ once-a-quarter), Archive (cheapest at rest, for rarely-read retention). All classes have millisecond first-byte latency — the colder classes differ by price, minimum storage duration, and retrieval cost, not by a rehydration wait. Autoclass moves objects between classes automatically by access.
- Location type (the durability/geo knob) — regional (single region), dual-region, or multi-region (geo-redundant); GCS handles intra-location durability internally.
- Cloud CDN / Media CDN — CDN in front of a GCS origin (→
content-delivery).
When to pick which
Standard for served objects; Nearline/Coldline for backups by access frequency; Archive for long-term retention. Regional for lowest cost/latency; multi-region for geo-redundancy and availability across a region loss. Autoclass when access patterns are unpredictable.
Limits / things that bite (verify against current docs)
- Resumable upload is the recommended path for large objects and flaky links (session-URI based, query-and-resume); multipart (XML API) also exists for S3-compatibility.
- Colder classes carry minimum storage durations and per-GB retrieval fees — early deletes and tiny churny objects can cost more than Standard despite the lower storage rate.
- Egress to the internet is billed (and inter-region transfer); Cloud CDN cuts egress + origin load.
- Request rate auto-scales but ramps; very spiky write bursts to a fresh bucket can need a warm-up.
- Unlike most stores, even Archive is instant-access — do not assume an archive "thaw" step.
Provider-specific trade-offs
- Integrates with IAM, Pub/Sub notifications (object events →
messaging-streaming), and BigQuery/Dataflow for data-lake reads; convenient but GCP-specific. - Multi-region buckets simplify geo-redundancy without app-managed replication, at higher $/GB.
Pitfalls
- Deleting Nearline/Coldline/Archive objects before the minimum duration → early-deletion charges.
- No lifecycle rule to expire noncurrent versions or tier down → silent storage growth.
- Serving popular objects from origin without Cloud CDN → egress blowup.
- Over-broad or long-lived signed URLs — scope to object + method, keep expiry short.