An object storage service on Google Cloud Platform (GCP). Unstructured data is stored as objects inside buckets rather than as files on a provisioned disk, so applications can write, read, and list arbitrarily large datasets without managing servers or volumes.
Each object is an immutable blob with a name, optional metadata, and the bytes themselves. Typical contents include media libraries, data-lake files, backups, static website assets, machine learning (ML) training sets, and logs. Google publishes 99.999999999% annual durability for stored objects, and every storage class is online, with millisecond retrieval rather than a tape restore.
Buckets, Objects, and Locations
Buckets are the top-level containers. They cannot nest, and the name is globally unique. Creating a bucket also chooses a location that is not renamed in place:
- Region: a single geographic area, usually the lowest latency to nearby compute.
- Dual-region: two regions. Applications keep reading if one region is impaired, with a Recovery Time Objective (RTO) of zero in Google’s dual-region design.
- Multi-region: a large geographic area such as the United States or Europe, aimed at readers in many places.
Object names may contain slashes, which the console presents as folders. The default keyspace is still flat unless hierarchical namespace is enabled on the bucket. Objects are addressed with gs:// Universal Resource Identifiers (URIs). The same objects are also reachable as ordinary Hypertext Transfer Protocol (Secure) (HTTPS) Uniform Resource Locators (URLs).
Storage Classes
Every object has a storage class that sets the price of keeping the bytes, whether reads are billed as retrieval, and how long the object must remain in that class. A bucket can set a default class. Autoclass can move objects between classes from last-access time, and Object Lifecycle Management can transition or delete on age, prefix, or version count.
| Storage Class | API Name | Minimum Duration | Retrieval Fees | Typical Availability |
| Rapid | RAPID | None | None | >99.95% in zones |
| Standard | STANDARD | None | None | >99.99% in multi-regions and dual-regions; 99.99% in regions |
| Nearline | NEARLINE | 30 days | Yes | 99.95% in multi-regions and dual-regions; 99.9% in regions |
| Coldline | COLDLINE | 90 days | Yes | Same as Nearline |
| Archive | ARCHIVE | 365 days | Yes | Same as Nearline |
Rapid is the zonal class for high-throughput artificial intelligence (AI) and analytics jobs, often paired with Rapid Cache, a read cache backed by solid-state drives (SSD). Standard is the default for data that is read often. Nearline, Coldline, and Archive cut storage price in exchange for a minimum duration and retrieval fees, while remaining online through the same Application Programming Interface (API).
Access, APIs, and Security
Clients use the JavaScript Object Notation (JSON) API, an eXtensible Markup Language (XML) API that many Simple Storage Service (S3) tools can target, language Software Developer Kits (SDKs), the console, and the gcloud or gsutil Command Line Interface (CLI). Access is granted with Identity and Access Management (IAM) at the project or bucket. Uniform bucket-level access is the recommended model because it disables per-object Access Control Lists (ACLs).
Temporary access uses signed URLs, commonly Hash-based Message Authentication Code (HMAC) V4 signatures. Encryption at rest is on by default. Teams can bring customer-managed keys through Cloud Key Management Service or supply their own keys. Retention policies, bucket lock, object holds, and soft delete cover regulated retention and accidental deletion. Browser apps from other origins need Cross-Origin Resource Sharing (CORS) headers. A Content Delivery Network (CDN) can sit in front of public objects for media delivery.
Integrations and Common Workloads
Once objects are in a bucket, the rest of Google Cloud can read them in place:
- Analytics: BigQuery can query files in place (Comma-Separated Values (CSV), JSON, Avro, Parquet, ORC).
- Compute and training: Google Kubernetes Engine (GKE), Compute Engine, Tensor Processing Unit (TPU), and Graphics Processing Unit (GPU) jobs read training data directly. Cloud Storage FUSE mounts a bucket with filesystem semantics when a trainer expects local paths.
- Events: Publish-Subscribe (Pub/Sub) notifications fire when objects are created, updated, or deleted.
- Protection: object versioning keeps noncurrent generations. Dual-region turbo replication offers a 15-minute Recovery Point Objective (RPO) Service-Level Agreement (SLA).
- Migration: Storage Transfer Service, Transfer Appliance, and Cloud FTP (via Secure File Transfer Protocol (SFTP)) move data in.
The data model matches Amazon S3 on Amazon Web Services (AWS): a bucket, an object, and a class. Google-specific pieces are Rapid and Autoclass, dual-region turbo replication, and the direct path into BigQuery and GKE.
Additional Acronyms for GCS
- GCS - Geographic Coordinate System