Meta Rebuilds BLOB Storage to Cut GPU Stalls in AI Training

Meta has re-architected its exabyte-scale BLOB-storage stack to meet the demands of AI training, where GPU synchronization steps mean that any slow storage fetch stalls the entire cluster. The legacy system, built for web-scale HDD workloads, relied on multiple stateful metadata layers (name, volumes, container) that could add hundreds of milliseconds of latency per lookup — fine for traditional apps, but incompatible with the millisecond-level, bounded pMax latency AI pipelines require from flash storage.

The new design collapses metadata into a single flat schema backed by ZippyDB, enabling O(1) path resolution instead of multi-hop lookups. It also removes the dataplane proxy in favor of a fat client SDK that streams bytes directly from Tectonic storage servers to GPU hosts, cutting both latency and the power overhead of proxying traffic — a critical consideration as datacenters become power- rather than space-constrained. Finally, the stack can now be deployed regionally, colocated with GPU clusters, rather than only as a global service.

Together these changes let Meta match Tectonic's raw performance with effectively zero storage-layer overhead, directly improving GPU utilization and reducing the data-movement friction that slows AI research iteration across geo-distributed compute.