1.4 KiB
Fanyu high-capacity resource profile
Enable DROBOTICFLOW_RESOURCE_PROFILE=fanyu-high-capacity-v1 on the Worker with
K8S_NAMESPACE=droboticflow-fanyu. Other namespaces are unaffected.
Effective limits are raised after platform, operator, and released workflow node overrides are resolved. Historical releases and Runs are not rewritten. Regular operators receive at least 8 CPU / 16 GiB; video, rectification, blur, slice and delivery operators 16 CPU / 32 GiB; AI, GPU, model training and pose operators 16 CPU / 64 GiB. WiLoR and HeadPose inference sidecars use the same profile. Larger limits and reservations are preserved. GPU allocations, resource requests, shared memory, concurrent Capture limits and replica counts do not change. Limits are ceilings, not dedicated capacity; concurrent jobs can still contend.
API limits: 8 CPU / 16 GiB. Worker limits: 4 CPU / 8 GiB.
Apply the documents in k8s/fanyu-high-capacity.resources.yaml using strategic
merge patches to existing deployments, not as replacement manifests. Preserve
any already-higher limits. Scaling dormant model services is not required.
New operator Jobs receive the profile; existing Jobs are not restarted or mutated. The interrupted slice task is not automatically retried by changing resources. Removing the Worker profile environment variable restores normal precedence for future Jobs; deployment limits can be restored from the pre-change backup.