Spotify Backstage: The Catalogue as a Product

The problem

At Spotify's scale the questions that consume engineering time are mundane and unanswerable: which service does this, who owns it, where is its documentation, what does it depend on, how do I create a new one correctly, is it healthy.

Each of these had an answer somewhere — a wiki, a spreadsheet, a dashboard, a colleague — and the aggregate effect was that finding out took longer than doing the work. That cost scales with the number of services and the number of engineers, so it gets worse precisely as the organisation grows.

What they did

Backstage, open-sourced in 2020 and later donated to the CNCF, is a developer portal built around a software catalogue. Every service, library, data pipeline and website is registered with its owner, its dependencies, its documentation and its operational state — and the portal is extensible through plugins so that CI status, monitoring, cost and security findings appear alongside.

Two features carry most of the value. Software templates scaffold a new component that is already wired into the organisation's pipeline, telemetry and conventions — so a service starts correct rather than accumulating correctness through review comments. And documentation lives with the code and is rendered in the portal, which is what keeps it current.

The property that determines success

Whether the catalogue data is derived or declared. A catalogue populated by hand is accurate on the day it is filled in and misleading within a quarter — which is worse than nothing, because people trust it.

One that ingests ownership from repository metadata, dependencies from build manifests and traces, versions from the deployment system and health from monitoring stays true without anyone maintaining it. Ownership decays fastest, because reorganisations happen more often than services are retired, and it must be tied to a group identity that survives individuals leaving.

The transferable lesson

Usage signals turn out to be the most valuable metadata and the least anticipated. Knowing a service is called daily by forty other services is a stronger trust signal than any documentation field; knowing something has not been called in a year is what makes decommissioning possible.

And the adoption test: engineers open it because it is faster than asking a colleague. A catalogue built for compliance reporting is populated, mandated and unused.