Horizontal scaling

Run the API as stateless replicas behind a load balancer. Configure STORAGE_DRIVER=s3 so every node shares cache and batch objects, and set CACHE_DISTRIBUTED_LOCK_ENABLED=true with the same Redis REDIS_URL on every replica to coalesce concurrent cold-cache transforms. Disk storage is local to one node. Redis lock failures fail open; BullMQ also uses Redis when its queue is enabled.

The full horizontal scaling architecture and Mermaid diagram covers lock behavior, Redis/S3 operations, consistent-hash routing, HPA, and capacity limits. Its load-test report includes commands for 1k, 5k, and 10k RPS. Measured figures are intentionally unreported until those workloads run on a multi-node deployment.

Enable the Helm HPA with autoscaling.enabled=true; install Metrics Server and tune CPU/memory requests and limits using the documented load procedure. See Deployment for the base chart and Kubernetes setup.