GrowthBook

GrowthBook inverts the usual experimentation architecture — it does not ingest your events. SDKs download a feature-definition payload, cache it, and evaluate flags locally; exposures get logged through whatever pipeline you already run (Segment, RudderStack, direct to warehouse). GrowthBook then connects to your warehouse and computes experiment results with SQL against your own tables. It runs read-only SELECT queries and only aggregated statistics leave the warehouse — raw user data stays put, which is why privacy-sensitive teams like the model. Metrics are defined once against warehouse data, so experiment analysis reconciles with your reporting by construction, and every result exposes the underlying SQL (exportable as a Jupyter notebook) rather than a black-box number.

The warehouse support is broad: BigQuery, Snowflake, ClickHouse, Databricks, Redshift, Postgres, MySQL/MariaDB, MS SQL Server, Athena, Presto/Trino, Vertica, and Mixpanel as a non-SQL source. On Cloud you whitelist a single GrowthBook IP for the warehouse connection; the docs recommend a read-only database user scoped to just the tables it needs.

Statistics are a genuine strength. The default engine is Bayesian (reporting “chance to win” with an uninformative prior by default; proper priors are an org-level opt-in), with a frequentist engine that adds sequential testing and CUPED-style regression adjustment for variance reduction. Built-in data quality checks — sample ratio mismatch detection, multiple-exposure alerts, guardrail metrics — catch the classic ways experiments silently go wrong.

Self-hosting notes: the whole app (Next.js front-end, Express API, Python stats engine) ships as one Docker image, but it requires MongoDB (or a compatible service like DocumentDB or Cosmos DB) for metadata — a dependency to plan for if your ops stack has no document store. A Helm chart exists for Kubernetes. SDK coverage is wide: JavaScript/React/Vue, iOS, Android, Flutter, React Native, plus server-side Node, Python, Ruby, PHP, Java, Go, C#, Rust, Elixir, and edge SDKs for Cloudflare Workers, Fastly Compute, and Lambda@Edge.

Gotchas: the visual editor, multi-armed bandits, and safe rollouts are Pro-tier features on Cloud, not part of the free Starter plan; Cloud plans also meter CDN requests for flag delivery. And because GrowthBook supplies no event pipeline, assignment/exposure tracking quality is entirely on you — bad instrumentation produces confidently wrong experiment results.

Use it when you already have trustworthy event data in a warehouse and want flags plus experiment analysis on top of it, self-hosted or cloud, without per-event vendor pricing. Look past it when you have no event pipeline to build on, or marketers need a fully bundled suite with built-in behavioral analytics rather than a warehouse-first workflow.