Problem
Metric definitions at a large company drift apart: the same metric lives in several datasets and pipelines, and teams follow different rules. At Netflix this hurt most in the experimentation platform. Adding a metric needed business knowledge, familiarity with experimentation conventions and complex engineering work, so it took weeks. That delay was prohibitive for data scientists and analytics engineers who wanted to onboard new metrics.
Example from the article: onboarding one new metric to the Netflix experimentation platform could take weeks before DataJunction.
Solution in brief
Netflix adopted DataJunction (DJ), an open-source semantic layer that parses SQL into a connected graph of metrics, dimensions and tables, and generates SQL back for consumers. A small team built the core first: the graph engine plus an API. Cubes, which bundle metrics and dimensions, became the way clients consume it.
Questions this case answers
- How to keep a single governed definition of each metric when many teams and tools compute it differently?
- How to let analysts add new metrics to a shared platform without deep engineering knowledge of its internals?
- How to define metrics once so that experiments and dashboards use the same definitions?
- How to cut the time needed to add a new metric to an experimentation platform from weeks to something an analyst can do alone?
Tools and methods
- DataJunction (DJ) — open-source semantic layer: stores definitions as a graph, parses and generates SQL
- Cubes — bundles of metrics and dimensions served to clients as definitions and as fast data access
- Apache Druid — target where DJ materializes cubes for a fast analysis tool
- Apache Superset — example of an external tool that can show DJ metrics with a link back to the definition
Google or a one-time email code. No password.