Build on the spot-preemption drain so that when a backend withdraws previously-advertised headroom (see the capacity-advertisement issue), any guest workloads scheduled into that headroom are drained within a bounded window (default 120s) rather than hard-killed, and the control plane is notified to reschedule.
Acceptance: withdraw triggers a bounded drain; workloads exceeding the window are reported; no host wedge on repeated advertise/withdraw cycles.
Part of a control-plane capacity-pooling effort (internal tracking: FootprintAI/Containarium-cloud#491).
Build on the spot-preemption drain so that when a backend withdraws previously-advertised headroom (see the capacity-advertisement issue), any guest workloads scheduled into that headroom are drained within a bounded window (default 120s) rather than hard-killed, and the control plane is notified to reschedule.
Acceptance: withdraw triggers a bounded drain; workloads exceeding the window are reported; no host wedge on repeated advertise/withdraw cycles.
Part of a control-plane capacity-pooling effort (internal tracking: FootprintAI/Containarium-cloud#491).