Meta redesigned its blob-storage architecture to address performance bottlenecks in AI workloads by implementing a unified metadata schema backed by ZippyDB, eliminating dataplane proxies in favor of fat client SDKs, and deploying regional storage systems colocated with GPUs. The redesign introduced tiered caching, using host memory and flash as L1/L2 caches alongside regional flash as an L3 tier, combined with prefetching APIs that let researchers access data globally without regional ingestion delays. These changes reduced GPU stalls caused by storage latency and cut data ingestion times, letting researchers iterate faster on model development.
