Keeping work available
Cordon. Cordoning makes a node ineligible for new placements but leaves its existing runtimes alone. They may finish naturally, but the cordon itself neither moves them nor terminates the node. The flag survives a norsk-mgr restart. Uncordoning makes the node eligible for placement again.
Drain. Draining is an active, node-wide operation: the Manager cordons the node, starts a migration for every runtime on it and terminates the node once it is empty. If a runtime cannot be moved before the timeout, an ordinary drain ends partially and leaves it running on the cordoned node; a forced drain terminates the node and lets normal failure recovery handle anything left behind. If you want a node merely to receive no new work and empty as its current work completes, cordon it rather than draining it.
Migration. A migration moves one runtime from its current node to another suitable node. Its strategy comes from the product template: stopThenStart stops the source before starting the prepared target and therefore has a gap; overlap briefly runs both; handover uses product cooperation to transfer state and coordinate the cutover. A drain runs this same migration machinery for every runtime on the node.
Eviction and termination. Both operations take a node down immediately and are normally refused while it hosts runtimes unless they are forced. Eviction adds a durable cordon first, so a worker that returns under the same identity remains excluded from placement. A forced eviction drops the runtimes with the node; eligible runtimes are then recovered elsewhere with a gap. Drain is the graceful alternative.
Hot spare. A pool’s hot-spare policy asks the Manager to keep ready, idle capacity available, optionally warmed or reserved for particular work. A live spare is an ordinary node in a spare state, not a different kind of node. A local server may be designated as spare, but the Manager cannot replace the physical machine if it disappears. A migration lands on a spare held for other work only when the operator overrides reservations for it (the Migrate dialog’s “Override reservations”, or runtime migrate --override-reservation); with that set, the spare can also be picked as the migration’s target.
Node service. A node service is intended to be a controller-managed, long-running service placed using pools and capabilities rather than owned by one production. Its definition currently comes from Manager configuration and is read-only through the operator API. This part of the model remains experimental and under design.