Designing a Feature Flag System

Feature flags decouple code deployment from feature activation, enabling safe and gradual rollouts.

Types of Feature Flags

Release toggle:    Enable/disable in-development features
Kill switch:       Emergency off switch for problematic features
Experiment toggle: A/B test variants
Permission toggle: Feature only for specific users/plans
Ops toggle:        Runtime behavior tuning

Data Model

{
  "flag_key": "new_checkout_flow",
  "enabled": true,
  "rules": [
    {
      "type": "percentage",
      "percentage": 10
    },
    {
      "type": "user_ids",
      "ids": ["user_123", "user_456"]
    },
    {
      "type": "attribute",
      "attribute": "plan",
      "operator": "equals",
      "value": "enterprise"
    }
  ],
  "default": false
}

Evaluation Algorithm

def evaluate(flag_key, user):
  flag = get_flag(flag_key)
  if not flag.enabled:
    return flag.default

  for rule in flag.rules:
    if rule.type == 'user_ids':
      if user.id in rule.ids:
        return True

    elif rule.type == 'percentage':
      # Deterministic hash: same user always gets same result
      hash_val = hash(f"{flag_key}:{user.id}") % 100
      if hash_val < rule.percentage:
        return True

  return flag.default

Architecture

Flag Config Store (PostgreSQL)
    ↓
Flag Service API
    ↓
SDK (in each service) — evaluates locally
    ↓
Cache (Redis) — configs synced every 30s
    ↓
Application code

Performance

Flag evaluation must be <1ms — sync flags to in-process cache:

Startup: Load all flags into memory
Refresh: Poll API every 30s for changes
Fallback: Last known config if API unavailable

Experimentation Integration

Flag evaluation event logged:
{ flag: 'new_checkout', variant: true, user_id: 123 }

→ Joined with business metrics (conversion, revenue)
→ Statistical significance calculated
→ Winning variant identified
→ Rollout to 100%

Conclusion

Feature flags are the enabler of continuous deployment. Local evaluation with periodic sync ensures <1ms overhead and resilience to central service failure.