Why must services be stateless to scale?

Horizontal scaling works when any instance can serve any request and instances can be killed and replaced without anyone noticing. If an instance holds state — a user's session in memory, an uploaded file on local disk, a cache that other instances do not have — then requests must be routed to that specific instance, it cannot be replaced without losing something, and rolling deploys log users out. The classic incident is autoscaling plus round-robin plus in-memory sessions: random logouts.

State is externalised: sessions in Redis with a TTL or in a signed token the client carries; uploads to object storage; shared caches in Redis rather than per-process. The cost is real: an extra network hop per request (Redis at ~1 ms versus a memory read), and Redis becomes a dependency whose outage logs everyone out at once, so it needs its own availability story.

Sticky sessions are a workaround, not a solution: they fight the load balancer (uneven load), break when an instance dies, and complicate deploys. They are acceptable temporarily while sessions are being moved out.