You need to A/B test a new product-recommendation algorithm behind a global HTTP(S) load balancer and randomize users. How should you route traffic?

Splitting traffic between versions using weights is the most effective method for A/B testing with randomization. This approach allows you to direct a controlled percentage of user requests to the new algorithm while the majority still uses the old one, enabling direct comparison of performance metrics. This method inherently randomizes users based on the load balancer's distribution. Enabling a feature flag on a single instance is not suitable for global A/B testing as it limits the test scope and doesn't provide a randomized, representative sample across all users. Mirroring traffic is primarily for testing in a production-like environment without impacting live users, not for actively routing user traffic for A/B comparison. HTTP header-based routing is useful for targeted testing (e.g., internal users) but doesn't provide the necessary randomization for a true A/B test of a global user base.