Metric Semantic Layer: How Lyft Governs and Scales Key Data Definitions

Rohit Channe, Simran Mirchandani10 Jun 2026

Problem

As Lyft grew, products launched and people moved between teams, so different teams began to use different definitions of the same metric. Without version control or a shared standard, outdated definitions crept into decision-making. To find the right definition, people searched codebases or asked colleagues who remembered it.

Example from the article: asked what "Metric ABC" meant, a team would answer differently depending on the context and application it had in mind.

Solution in brief

Lyft built a Metric Semantic Layer (MSL): a versioned Python package that stores each shared metric once, as a YAML config with a plain-English description and a SQL template. Only "Golden Metrics" (at least two distinct uses) are admitted, and each has a business owner and an operational owner who must both approve changes. A Python API, a data catalog, a self-service UI and an MCP server for AI agents read the same configs.

Questions this case answers
  • How to keep one governed definition of each business metric when many teams and tools reimplement it independently and definitions drift
  • How to assign ownership of shared metric definitions so they stay current through team rotations and attrition
  • How to give LLM agents a deterministic source of metric logic so natural-language questions return consistent SQL
  • How to give every team one authoritative definition of each product metric and keep it in sync with dashboards, ML models and pipelines?

Tools and methods

  • YAML metric configs — owners, data sources, dimensions, granularities, metric expressions
  • Jinja SQL templates — one SQL block shared across definitions (DRY, lower learning curve than other templating frameworks)
  • Golden Metrics rule and two-owner model — what enters the package and who maintains it
  • Python package with get_sql — ready SQL for a metric, dimensions and date range
  • Amundsen — catalog with definitions, owners and data quality information
  • Metrics MCP server — search and query tools for AI agents, with authorization and evaluation
Continue reading — it's free

Google or a one-time email code. No password.