Feature Flags & Progressive Delivery

A runtime toggle that controls whether a feature is active for a user or group; enables deploying code separately from releasing it.

"We shipped the new checkout behind a feature flag — it was deployed to production for two weeks of testing before we enabled it for any users."

The stages a feature flag passes through: create, enable for internal testing, progressive rollout, full release, and then retire the flag from the codebase.

"We track all flags in a spreadsheet with a retire-by date — flags that stay past their date are treated as technical debt in the sprint."

Condition that determines which users see a specific flag variation: by user ID, organisation, percentage, geographic region, or custom attribute.

"The targeting rule enables the new dashboard only for users whose plan attribute equals enterprise — free tier users see the existing version."

Gradually enabling a feature for an increasing percentage of users while monitoring error rates and metrics before full release.

"We rolled out the new search engine to 1%, then 10%, then 50% over three days — each step was gated on p99 latency staying below 200ms."

Routing a small percentage of production traffic to a new service version while the rest uses the stable version; named after canaries in coal mines.

"The canary release sent 5% of requests to v2 — when error rate was indistinguishable from v1 after 30 minutes, we shifted all traffic."

Running two identical production environments; traffic switches atomically from the blue (current) to the green (new) environment with instant rollback capability.

"Blue-green deployment gave us zero-downtime releases — switching traffic took 10 seconds and rolling back was equally instant by swapping the load balancer target."

Rolling out changes through concentric groups of increasing size: ring 0 is internal users, ring 1 is beta users, ring 2 is the broad release.

"Microsoft uses ring deployment for Windows updates — issues surface in ring 0 and 1 before reaching the millions of devices in the general release ring."

A feature flag configured to instantly disable a feature in production without requiring a new deployment.

"When the new payment processor started declining legitimate cards, we flipped the kill switch and restored the old processor within 30 seconds."

Technical debt from feature flags that were never cleaned up after their rollout completed; stale flags increase code complexity and cognitive overhead.

"A code review revealed a flag from 18 months ago still wrapping the new onboarding flow — cleaning it up removed 300 lines of dead code and the old flow."

Deploying new code behind a flag that is off for all real users; allows load testing and infrastructure validation in production without user exposure.

"We dark-launched the new recommendation engine by duplicating real requests to it and logging the results — no user saw the output until we were confident in quality."

A branching strategy where all developers integrate into a single main branch at least daily; feature flags allow incomplete work to coexist safely.

"Trunk-based development with feature flags replaced our long-lived feature branches — merges are trivial and the main branch is always deployable."

Popular commercial feature flag management platform offering targeting rules, A/B experimentation, and real-time flag updates via a streaming connection.

"LaunchDarkly's streaming SDK means flag changes reach all connected clients in under 200ms — no polling delays between toggling a flag and seeing the effect."

Open-source feature flag server available as a self-hosted option; supports multiple strategies including gradual rollout and user-based targeting.

"We self-host Unleash so feature flag data never leaves our infrastructure — a compliance requirement for our enterprise customers."

The process of determining which variation a specific user sees by evaluating targeting rules in order until a match is found.

"Flag evaluation runs client-side in the SDK with a locally cached ruleset — there is no network request per evaluation, so it adds zero latency."

An automated condition — such as an error rate spike or latency threshold breach — that disables a flag or reverts to the previous deployment state.

"We configured an automated rollback trigger: if error rate exceeds 1% within 5 minutes of a canary shift, the flag reverts to the stable variation automatically."