How do you analyze feature flag and experiment data in Metabase?
Feature flag and experimentation tools record what changed for users and when: which flag was turned on, for whom, and what the experiment showed. To analyze them in Metabase, you sync each tool into a database and model flags, flag changes, and experiment results. Metabase has no native connector for these tools, so the sync comes first. This is a subcategory of software delivery integrations.
TL;DR — Model flags (or gates), flag changes, and experiment results. Group raw toggles into meaningful changes before trending them, and report the experiment decisions the tool made instead of recomputing its statistics.
Which tools does this cover?
This pattern applies to feature flag and experimentation tools, including:
- LaunchDarkly — flags, flag changes, and experiment results
- Statsig — experiments, gate changes, and Pulse results
The same model fits other flag platforms’ audit logs and experiment exports.
What is the shared feature flag data model?
| Concept | LaunchDarkly term | Statsig term | Used for |
|---|---|---|---|
| Flag | Flag | Feature gate | Inventory, flag debt |
| Flag change | Audit-log entry | Gate change | Rollout activity, change failure |
| Environment | Environment | Environment | Keeping production separate |
| Experiment | Experiment | Experiment | Velocity, decisions |
| Result | Experiment results | Pulse results | Lifts by metric and variant |
| Exposure | Exposures | Exposures | Reach |
Key flag changes on flag, environment, and timestamp so they line up with deploys and incidents on the same timeline.
How do you connect feature flag tools to Metabase?
- Pipeline route — sync flags, audit-log entries, and experiment results into a database on a schedule with a connector or the tool’s API. See our guide to building a data pipeline.
- MCP + CLI route — pull a scoped slice through the vendor’s MCP server, save it as CSV, and load it with
mb upload csv. Good for exploration; it keeps no history.
Per-tool setup lives on each guide above.
What can you analyze across these tools?
- Rollout activity — meaningful flag changes per day, by environment and team
- Stale-flag debt — temporary flags still live past their planned lifespan
- Experiment velocity and decisions — experiments started, concluded, and shipped
- Deployment frequency — read next to rollouts, since flags decouple deploy from release
- Change failure rate — count rollouts that caused incidents, not only deploys
- Error rate — errors before and after each rollout
- Lead time for changes — commit to deploy, with rollout as a separate step
- Feature adoption rate — usage of a feature once its flag is on
Which dashboards should you build?
- Feature flag rollout — flag changes, stale flags, and rollouts next to deploys
- Release quality — failed changes and incidents per release, rollouts included
- Software delivery dashboard — the delivery view flags plug into
Common mistakes
Trending raw toggle counts. → One rollout can be fifty micro-adjustments. Group audit events into meaningful changes (per flag per environment per day) before trending.
Declaring winners from exposure rollups. → Statistical judgment lives in the experimentation tool. Report its decisions and lifts; don’t recompute significance from daily aggregates.
Syncing evaluation events. → Evaluation events are high-volume and belong in the flag tool. Land audit events and result rollups.
Ignoring temporary flags past their lifespan. → Old temporary flags are dead code paths waiting to break. Track them as flag debt with an owner.
Related
Analytics
Integrations
FAQ
Does Metabase connect natively to LaunchDarkly or Statsig?
No. Metabase reads SQL databases and warehouses, not SaaS APIs. Sync flags, audit logs, and experiment results into a database with a connector or the tool’s API, then connect Metabase to that database. For quick exploration, pull a slice through the vendor’s MCP server, save it as CSV, and load it with
mb upload csv.Should I compute experiment significance in Metabase?
No. Statistical judgment stays in the experimentation tool, which has the raw exposures and the right methods. In Metabase, report decisions, lifts by metric, and experiment velocity.
Why analyze flag data next to delivery data?
A rollout is a production change, just like a deploy. Join flag changes to deploys and incidents to get an honest change failure rate, and to application errors to see the error rate around each rollout. See the software delivery overview for the deploy side of the model.