Skip to content

SYSTEM Cited by 1 source

Tectonic (Meta block storage)

Definition

Tectonic is Meta's regional, multi-tenant, horizontally scalable foundational block-storage layer upon which all higher-level storage APIs (object storage, file systems, block devices) are built. It is the bottom of Meta's storage stack.

Key Characteristics

  • Regional deployment: operates within a single region (not globally replicated by default)
  • Multi-tenant: serves all Meta products — Facebook, Instagram, Reality Labs, Meta AI, Ads, Data Warehouse, internal Databases
  • High durability and availability via erasure coding
  • Media tiering: supports both HDD and flash, with smart placement of hot/cold/warm data for efficient I/O utilization across tenants
  • Hundreds of exabyte-scale clusters in production

Role in the AI Storage Stack

The 2026-07-01 BLOB-storage re-architecture post confirms Tectonic as the terminal data layer:

  • The new BLOB-storage SDK embeds a Tectonic BlockClient directly, streaming data from Tectonic block servers without an intermediate proxy
  • The getReadPlan() API resolves paths to (blockId, offset, size) tuples — these are Tectonic-native addresses
  • The design goal of the new architecture was "adding zero overhead on top of Tectonic"
  • The tiered cache hierarchy (L1 memory → L2 flash → L3 regional flash) ultimately fetches from Tectonic when cache misses cascade

Prior Art

An earlier @Scale talk titled "Training Llama: A Storage Perspective" described training Llama directly over Tectonic's block layer by exposing an NFS-like FileSystem interface on top of it. While this architecture continues in use, modern training has migrated to the BLOB-storage interface for unified data-lake access and higher performance.

(Source: sources/2026-07-01-meta-ai-storage-blueprint-at-scale)

Seen in

Last updated · 567 distilled / 1,685 read