Changed the autocoder-engine namespace memory quota to enforce memory admission through requests.memory instead of limits.memory.
The namespace was consuming 62592Mi of limits.memory quota because seven idle Codex worker pods each declared an 8Gi memory limit. Actual runtime memory usage was low, but the summed limits left too little quota headroom for rollouts.
limits.memory: 64Gi from the autocoder-engine-quota ResourceQuota.requests.memory: 16Gi as the namespace memory admission cap.AutoCoder/AutoCoder-Engine commit 42d539f (Use request-based memory quota).After applying the change, kubectl describe resourcequota autocoder-engine-quota -n autocoder-engine showed:
requests.memory 8736Mi 16Gi
limits.memory was no longer present in the quota hard/used values.
This enforces admitted memory requests, not a hard shared runtime memory cap. Individual pod memory limits still apply where configured, and node memory pressure/OOM behavior remains responsible for simultaneous real memory spikes beyond node capacity.