dify/docker/kubernetes/dev-builder-resources.md

Dev API resource profile

dev-builder-resources.yaml records the reviewed four-role memory profile for ENG-1311 in the Dify repository. It is an environment-specific Helm values overlay, not a general default or evidence that the current Helm release uses it. The API affinity is also consumed by the WebSocket chart template.

The worker reservation is 2 GiB, above its observed 1.63 GiB working-set peak; the 3 GiB limit leaves room for initialization and task execution. Other reservations cover observed steady-state usage. API's existing 2 GiB limit is preserved, not certified as sufficient for arbitrary load. Requests improve scheduling and eviction priority; they do not reduce actual memory usage. Preferred anti-affinity encourages distributing these four roles and does not guarantee placement or capacity. No CPU limit is introduced.

Applying the profile

1. Capture current Deployment UID, full template, replica count, referenced configuration hashes and images for the four named roles. Record the database schema and existing unknown operation identities without reading arguments or credential values. Check all-namespace Pod requests, init containers, taints, architecture, readiness, actual memory use and placement capacity.
2. Derive candidate templates from the captured live templates. Change only the target container's resources and the reviewed affinity. Reject any other structural difference. Preserve existing unrelated affinity restrictions.
3. Require a quiet Builder and worker, stop API/WebSocket/Beat producers, close the exact main worker queue definitions with individually acknowledged consumer controls, and drain active/reserved/scheduled/unacked work. Do not clear broker queues or replay unknown operations.
4. Stop the drained worker. With all four roles at zero, apply each candidate through JSON Patch tests of Deployment UID, complete old template and replica count. Server dry-run must pass first. Do not use a rolling surge on a memory-constrained node.
5. Start worker, WebSocket, API and Beat serially. At each step verify capacity, one Ready replacement, unchanged image/configuration, worker registered tasks and exact queue definitions. If scheduling or acknowledgement is unknown, inspect state before taking another action; do not blindly repeat mutations.
6. Verify /health, role source identity, no new OOM/eviction, ordinary queue operation, unchanged schema and historical unknown identities. Keep raw manifests and queue/task details private; publish only sanitized evidence.

Rollback uses the same admission/drain and UID/template checks and restores only the captured resources/affinity. It does not roll back images or schema.

Persistence boundary

The known Dev Helm record differs from approved live images and plugin configuration. A full helm upgrade --reuse-values can restore those stale values and is not a safe way to apply this resource fix. Reconcile the intended release baseline and include this overlay before a future full Helm upgrade; render and compare the complete release, preserving Trigger Worker image fallback and unrelated deployments. A resource-only live patch is an operational mitigation until that desired-state integration is complete. Keep ENG-1311 open for that integration and representative load observations.