Definition
Plain language
A queue that mostly serves jobs in arrival order but allows small reorderings to keep the system flowing.
As stated in the literature
In Forge, a scheduling discipline that approximates first-in-first-out order within a moving window of pending trajectories, balancing fairness and throughput on heavy-tailed agent-RL workloads.
Why it matters: It keeps RL training pipelines flowing when some trajectories take much longer than others, without letting any job starve indefinitely.
For example, among the 32 oldest pending rollouts, a windowed FIFO scheduler may pick whichever finishes fastest first instead of strictly the oldest one.
Heard on the show
“The first is windowed FIFO scheduling.”Episode 090 — How MiniMax-M2 Bets That Sparsity Plus Verifiable Rewards Can Match Frontier Agents