@Scale: Systems and Technologies Featured in Talks

ServiceRouter is Meta's service-mesh solution, in production since 2012, routing tens of billions of requests per second across millions of hosts and tens of thousands of services. It is integrated directly into Thrift, Meta's primary RPC transport, which allows it to provide out-of-the-box observability and logging for all RPC traffic, and it covers decentralised service discovery, load balancing and cross-region routing. Conveyor is Facebook's unified continuous deployment system, described as achieving over 100,000 deployments each week across a diverse range of services. Conveyor pipelines are composed of nodes — including Build, Canary and Push — through which each version of a service's code passes as a "bundle" containing a packaged binary.

• ServiceRouter: hyperscale service mesh, in production since 2012
• ServiceRouter integration with the Thrift RPC transport
• Conveyor: continuous deployment platform and the team supporting it
• Conveyor pipeline nodes: Build, Canary, Push
• Contbuild handles compilation and unit tests for Conveyor builds

MobileConfig is a cross-platform configuration-management system that developers use to manage configuration values and distribute them to mobile clients, in production since 2015. It uses a client-server pull model, fetching fresh values on application start and periodically during a session, and ties parameters to internal tools including the Gatekeeper feature-flagging platform and Quick Experiment. msys is a project initiated by Messenger engineers in 2018 to build a foundational messaging layer shared across Meta's family of apps. Threads, an app built by the Instagram team for sharing text updates and joining public conversations, is documented in a write-up of its iOS development, which began in early 2023 and led to launch on 5 July after five months of technical work.

• MobileConfig: strongly typed, cross-platform config parameters
• MobileConfig features include Emergency Push and Mobile Config Canary
• msys: shared mobile messaging platform started in 2018
• Threads for iOS: built on the Instagram app with a different app delegate
• Threads read-only web version shipped a month after the app launch

Everstore is the storage system into which content uploaded to Facebook is placed, specialising in Binary Large OBjects (BLOBs), with video data making up the majority of bytes stored. Semantic Replication is a technique described as saving over 30% of video storage by reducing the number of copies retained, exploiting properties such as multiple encodings per video and videos' temporal affinity. Virtual video files, together with FFUSE and FUSE file systems, provide an abstraction over the MP4 container format so that video bytes can be manipulated and transferred across data centres, connecting to FFmpeg for encoding. AV1, standardised by the Alliance for Open Media in 2018, is covered in the context of real-time communications, where comparison with an existing H.264-based solution shows BD-BR savings of around 30%.

• Everstore: BLOB storage system for uploaded content
• Semantic Replication: over 30% video storage saving
• Virtual video files, FFUSE and FUSE file systems
• AV1: standardised by AOMedia in 2018, applied to RTC video
• AV1 scalability modes: temporal, spatial and quality

TAO is a write-through cache on top of a sharded database that historically served the majority of Facebook's online social graph workload, exposing an object and association data model with simple query APIs. As products moved to more complex read queries, global secondary indexing services were built using asynchronous replication, which did not provide Read-Your-Writes consistency. FlightTracker is a set of APIs and services that manage consistency across the social graph, providing uniform Read-Your-Writes guarantees for both TAO and the global secondary indexes. It operates independently of the data stores, allowing services to be decoupled and simplified.

• TAO: write-through cache over a sharded database
• Object and association (nodes and edges) data model
• FlightTracker: uniform Read-Your-Writes guarantees
• Independent of data stores, preserving latency and availability

Owl addresses the distribution of large, hot content — executables, code artifacts, AI models and search indexes — to hosts within Meta's private cloud. The write-up frames the task along three dimensions of fanout, size and hotness, with objects ranging from around 1 MB to a few terabytes, and describes an architecture of peer libraries linked into client binaries and dedicated tracker services managing the control plane. BladeRunner is covered as a stateful service for real-time client-server interactions, holding stream state for PHP and delegating event processing to a PHP lambda function that responds with a set of actions.

• Owl components: peer libraries and trackers
• Owl requirements: fast, efficient and reliable distribution
• BladeRunner: stateful service holding stream state for PHP
• Distributed PHP lambda for real-time interactions

The Design Systems Platform (DSP) is described as a foundational layer that allows any design system to represent its foundational elements as design tokens. It is one of three areas of work intended to support the more than 20 design systems at Meta, alongside theming infrastructure and a library of "chameleon components" that support many design systems. Each design system is composed of foundational elements such as colours, fonts and spacing, UI components, and design principles.

• Design tokens defined, then synced into code and design
• Theming infrastructure for faster UI updates
• Chameleon components with variants
• Named design systems include FDS, IGDS, MDS and the WhatsApp Design System