Replicas and autoscaling: what changes?

A replica is a running copy of a service. Fixed replicas stay active together while a load balancer routes requests; they are not stored switched off.

Autoscaling adjusts the count between minimum and maximum limits. The current implementation adds one copy above the configured CPU threshold and removes one below the reduction threshold. Defaults are 75% and 25%, with a cooldown between decisions.

Every copy reserves CPU and memory from the purchased capacity. A plan’s replica ceiling is not free servers or unlimited resources.

Multiple copies need shared sessions and S3 uploads. Scheduled jobs need a leader or lock to prevent duplicate execution.

Copies on one host do not protect against that host failing. Distribution between the United States and Europe requires configured servers, routing and available capacity. Application replicas do not automatically configure database high availability.