Which load-balancing algorithm, and why?

Round robin assumes requests cost the same and backends are identical. It fails when request cost varies — one instance stuck with three heavy reports while the next gets cheap health pings — or when instances differ in size (weighted round robin fixes the second). Least connections adapts to cost variance by sending to the instance with the fewest in-flight requests; it fails when connection count does not reflect load, such as long-lived idle connections, and it needs the balancer to see all traffic to count accurately.

Consistent hashing routes by a key (user id, cache key) so the same key lands on the same instance. The reason is affinity: an in-process cache or a WebSocket state that would be useless if requests scattered. It distributes less evenly, hot keys create hot instances, and virtual nodes are needed to smooth the ring; when an instance leaves, only its keys move.

Health checks decide who is in the pool. Active checks hit an endpoint every few seconds; passive checks watch real responses. The danger is a check that depends on a shared dependency — if the database is slow, every instance fails its check simultaneously and the balancer removes everyone, turning a slow database into a total outage. Checks should verify the instance, not the world, and balancers should refuse to empty the pool below a floor.