Feature Flag

A feature flag wraps a code path in a conditional statement evaluated at runtime. The flag's state—on or off—is determined by a remote configuration service, a local config file, or a database query. This eliminates the need for a full redeployment to change functionality. The evaluation logic typically follows a pattern: if (featureFlag.isEnabled('new-checkout', user)) { showNewCheckout(); } else { showLegacyCheckout(); }. The flag check is designed to be extremely low-latency to avoid impacting application performance.

Modern feature flags go beyond simple on/off switches. They support granular targeting rules that determine flag evaluation based on user context. Common targeting attributes include:

• User ID or email domain: For internal dogfooding or beta lists.
• Geographic location: For region-specific launches.
• Subscription tier: To gate premium features.
• Custom attributes: Any property passed in the evaluation context. This allows for ring deployments, gradually expanding a feature from 1% to 100% of a user base.

Not all flags control user-facing features. Operational flags (or kill switches) govern system behavior and infrastructure. They act as circuit breakers to gracefully degrade service under high load. Examples include:

• Disabling a non-critical recommendation widget when database latency spikes.
• Switching from a primary payment gateway to a backup provider instantly.
• Reducing logging verbosity in production without a hotfix. These flags are often managed by Site Reliability Engineers (SREs) and are critical for high-availability architectures.

When combined with an analytics pipeline, a feature flag becomes an experiment flag. The flag service randomly assigns users into treatment and control cohorts and tracks their behavior against key metrics. This transforms a simple toggle into a rigorous A/B/n testing tool. The flag system must ensure consistent bucketing, meaning a user always sees the same variant during an experiment session. Statistical analysis is then performed on the collected data to determine if the new feature drives a statistically significant lift in conversion or engagement.

A feature flag is a temporary construct, not a permanent part of the codebase. Every flag must have a defined lifecycle: Creation, Rollout, Full Release, and Crucially, Removal. A flag that remains in the code long after a feature is permanently on becomes technical debt. It increases cyclomatic complexity, creates untestable code paths, and risks accidental deactivation. Mature teams enforce flag removal by setting expiration dates on flags and integrating automated cleanup tasks into their sprint cycles.

Feature flags are the primary enabler of trunk-based development for large teams. Instead of maintaining long-lived, divergent feature branches that result in painful merge conflicts, developers commit incomplete code directly to the main branch hidden behind a disabled flag. This practice ensures continuous integration of all work. The hidden code is deployed to production but remains inert, allowing teams to test in the real production environment without exposing unfinished features to end-users.

A randomized experimentation method where two versions of a variable are compared against each other to determine which performs better. Feature flags are the primary technical mechanism for delivering A/B tests in production.

• Flag-per-treatment: Each variant is gated behind a distinct flag rule
• Metrics integration: Flag evaluation events feed directly into experimentation platforms
• Example: Testing a new checkout flow by exposing 10% of users via a flag, measuring conversion lift against the control group

A deployment strategy that reduces risk by rolling out a new software version to a small subset of users before a full release. Feature flags enable instant canary rollback without redeployment.

• Traffic splitting: Route a percentage of requests to the new code path
• Health monitoring: Automatically disable the flag if error rates spike
• Example: Releasing a new search algorithm to 5% of users, monitoring latency and error budgets, then expanding to 100% over 24 hours

A source-control branching model where developers collaborate on code in a single branch called trunk or main. Feature flags are essential for this practice, allowing incomplete features to be merged without exposing them.

• Dark launching: Merge code early, keep it hidden behind a flag
• Continuous integration: Avoid long-lived feature branches that cause merge hell
• Example: A team merges a half-built recommendation engine daily, with the flag disabled in production until the UI and algorithm are complete

A language-specific library that integrates with a feature flag management service to evaluate flag rules at runtime. The SDK handles local caching, network resilience, and real-time updates.

• Evaluation context: Passes user attributes to determine targeting
• Fallback values: Defines safe defaults if the service is unreachable
• Example: A Node.js SDK evaluates isPremiumUser and region against flag rules to decide whether to show a beta dashboard, with a 10ms local evaluation latency

A special type of feature flag designed as an emergency circuit breaker for production systems. It allows operators to instantly disable a feature or code path without deployment when catastrophic failure is detected.

• Operational toggle: Separate from release flags, controlled by ops teams
• Circuit breaker pattern: Automatically trips on error threshold breaches
• Example: A payment processing flag that can be flipped globally to disable a new gateway integration if transaction failure rates exceed 1%

The conditional logic that determines which users see a feature based on user attributes, random percentages, or cohort membership. Targeting is the core intelligence layer of a feature flag system.

• Attribute-based: Target by email domain, plan tier, geography
• Percentage rollout: Randomly assign a consistent subset of users
• Example: A rule targeting beta_opt_in = true AND country = 'US' to expose a new analytics module only to opted-in American users