Sign in
Internal

Take one serving replica out of rotation (platform admin)

Drains this replica's pool member so it takes no new traffic while in-flight requests finish, and records a durable hold on the replica so the reconcile loop keeps it drained instead of re-activating it on its next pass. Keyed by vmId because that is the runtime-neutral replica identity: a pod replica is the same row shape as a VM replica. Returns the refreshed replica list.

post/v1/admin/inference/deployments/{id}/replicas/{vmId}/drain

Path parameters

idstring required
vmIdstring required

Response

Draining

affinityGuardUtilnumber double

prefix_affinity overflow bound; absent = the gateway default

policy'p2c_util' | 'prefix_affinity' required

The pool's member-selection policy (p2c_util when the pool row does not exist yet)

poolIdstring required

The deployment's endpoint pool id

strayMembersstring[]

Pool member hostnames no current replica owns (pending the reconcile loop's stray sweep). Usually empty.

Changes

Changed in 3 of the 19 revisions of this API.3