What is Cloud Storage Rapid?
Cloud Storage Rapid is a Google Cloud product family that accelerates object storage for data-intensive AI, ML, and analytics workloads. It bundles two offerings: Rapid Bucket, a dedicated zonal bucket using the new Rapid storage class built on Google’s Colossus storage system, and Rapid Cache, an SSD-backed zonal read cache (formerly Anywhere Cache). Both address the same bottleneck: accelerators such as GPUs and TPUs often wait for data because traditional object storage does not deliver enough throughput or has too much latency.
The approach behind Cloud Storage Rapid is to keep data in the same zone as the compute resources. This drives latency down to the sub-millisecond range and raises aggregate throughput significantly. Rapid Bucket delivers more than 15 TB/s of read throughput and up to 20 million requests per second from a single zonal bucket. As a result, expensive accelerators stay saturated and training and checkpointing times shrink.
Core Features
- Rapid Bucket with the Rapid storage class: Zonal object storage on Colossus with sub-millisecond latency, more than 15 TB/s aggregate read throughput, and up to 20 million requests per second from a single bucket.
- Streaming object appends: Applications can append data to existing objects in real time without rewriting them. This simplifies write patterns during training and checkpointing.
- Rapid Cache as a managed read cache: SSD-backed zonal cache with up to 2.5 TB/s aggregate throughput that works with existing regional, dual-region, and multi-region buckets with no code changes.
- Configurable caching: Caching in 2 MB chunks, ingest on read (default) or on write, strong consistency, and a TTL between 24 hours and 7 days.
Typical Use Cases
AI and ML training: When GPU or TPU clusters load large datasets, Rapid Bucket serves data at sub-millisecond latency from the same zone. This reduces blocked accelerator time and speeds up loading multi-modal training data.
Checkpointing: Streaming object appends and high throughput accelerate writing and restoring model checkpoints compared to traditional object storage.
Model loading and inference: Rapid Cache accelerates bursty read workloads such as loading models for inference from existing buckets, without requiring application changes.
Benefits
- Higher throughput and lower latency keep expensive accelerators saturated and shorten runtimes.
- Rapid Cache can sit in front of existing Cloud Storage buckets with no API or code changes.
- Consumption-based billing layered on Cloud Storage, with no upfront commitment.
Integration with innFactory
As a certified Google Cloud Partner, innFactory supports you with the adoption and operation of this service.
Typical Use Cases
Frequently Asked Questions
What is Cloud Storage Rapid?
Cloud Storage Rapid is a product family for high-performance object storage in Google Cloud. It comprises Rapid Bucket, a zonal bucket using the new Rapid storage class built on Colossus, and Rapid Cache, an SSD-backed zonal read cache. Both accelerate AI, ML, and analytics workloads that read from object storage.
When should I use Cloud Storage Rapid?
Use Rapid Bucket when you want data in the same zone as your accelerators and run training, checkpointing, or model serving at sub-millisecond latency. Use Rapid Cache when you have existing regional or multi-region buckets and need high read throughput for bursty workloads such as model loading, without changing code.
How much does Cloud Storage Rapid cost?
Billing is consumption-based and layered on Cloud Storage pricing. Data served from Rapid Cache can incur lower data transfer, retrieval, and operation fees than direct bucket access. Check the official Cloud Storage pricing page for current rates.
What limits and configurations apply to Rapid Cache?
Rapid Cache is a fully managed zonal cache that works with regional, dual-region, multi-region, and custom dual-region buckets with no API changes. A maximum of one cache is allowed per zone per bucket. Caching uses 2 MB chunks, with ingest on read (default) or on write and strong consistency. The TTL ranges from 24 hours to 7 days (default 24 hours).
