Skip to content

SYSTEM Cited by 1 source

Netflix Gutenberg

Gutenberg is Netflix's dataset pub/sub system — a versioned, dynamically-loadable, rollback-capable publication substrate for datasets consumed by multiple services. Canonical description is Netflix's "How Netflix Microservices Tackle Dataset Pub-Sub" post (not ingested on this wiki); the 2026-05-01 Model-Serving routing post names Gutenberg as the config-distribution substrate for Switchboard Rules.

This is a stub page; expand when the canonical Gutenberg post is ingested.

Role in model-serving routing (2026-05-01 post)

Quoting the post:

"Netflix's Gutenberg system provides an excellent ecosystem that enables a flexible pub-sub architecture, facilitating proper versioning, dynamic loading, easy rollbacks, and more. Both Switchboard and the Serving Cluster Host subscribe to the same Switchboard Rules configuration."

Load-bearing properties named:

  • Versioning — Switchboard Rules (A/B test configs, traffic splits, canary rules) get explicit versions; consumers can pin or follow the latest version.
  • Dynamic loading — consumers (Switchboard/Lightbulb + serving hosts) load new rule sets without restart.
  • Rollback — a bad rule set is rollback-able by reverting to a prior version; no code deploy, no fleet restart.

Why this matters for the serving platform

The Gutenberg substrate is what makes patterns/config-separated-from-code-via-pubsub work on the model-serving side. Researchers author Switchboard Rules as JavaScript → compile to JSON → publish via Gutenberg → Switchboard/Lightbulb + serving hosts subscribe. The experiment release cycle is decoupled from platform code deploys, which is the Switchboard Rules design's load-bearing property.

What's not yet canonicalised on this wiki

The canonical Netflix Gutenberg post (How Netflix Microservices Tackle Dataset Pub-Sub) is referenced by the 2026-05-01 Model-Serving post but not yet ingested. Topics still to canonicalise when that post lands:

  • Gutenberg's internal architecture (producer / subscriber / metadata store / versioning index).
  • Publication semantics (strict ordering? per-key? per-dataset?).
  • Race-condition flow for consumer sync (referenced by the 2026-05-01 post via diagram but not textually described).
  • Which other Netflix platforms use Gutenberg besides model-serving — the 2026-05-01 post hints at broader use but does not enumerate.

Seen in

Last updated · 445 distilled / 1,275 read