Airbnb Minerva: One Definition of a Metric

The problem

At a certain scale of analytics, every organisation encounters the same failure: two dashboards show different numbers for what appears to be the same metric, and both are defensible.

The cause is that the definition lives inside each individual query. One excludes cancelled bookings, another includes them until refund. One uses booking date, another check-in date. One filters internal test accounts, another does not. Nobody is wrong, and the organisation cannot answer its own question.

The cost is not the engineering time to reconcile. It is that executive attention goes into arguing about which number is right rather than into deciding anything.

What they did

Airbnb built Minerva as a central metrics platform: metrics and dimensions defined once, as version-controlled configuration, with the platform computing them consistently and every consuming tool querying through it.

The key architectural property is that the definition is upstream of the consumption tool. A metric defined inside a BI tool governs that tool only, and fails the moment a notebook or a second tool appears. A standalone layer with an API serves dashboards, notebooks, experimentation and applications from the same definition.

The trade-off

Central definition means a central bottleneck unless the contribution path is fast. If adding a metric requires a request that takes a quarter, analysts will define their own in SQL and the divergence returns — with the added insult that there is now a governance layer being bypassed.

There is also real computational cost: consistent computation frequently means pre-aggregation, which is storage and maintenance.

The transferable lesson

The blocker is organisational, not technical. Agreeing what "active user" or "revenue" means requires a decision from someone with the authority to make it, and the arguments that surface are ones the organisation has been avoiding for years. The platform is the easy half.

The practical sequencing that works: start with the small number of metrics that appear in executive reporting and regulatory submissions — where disagreement is most expensive — rather than attempting to define everything. And make the contribution path fast enough that using the layer is easier than bypassing it, because that, not policy, is what determines adoption.