Limitations

• Single-node batch jobs only. Multi-node, interactive, GPU, --constraint and reservation jobs are left to your rules.
• Pins only for plain jobs (one node, explicit memory, starts now; not --nodelist, --exclusive, held, or several tasks without -N 1). An array is not pinned at submission, only its tasks later, each on its own.
• "Room" = free now, not "starts next": higher-priority jobs may go first. Without always_place, jobs no node has room for are left to your rules.
• wait_for_room estimates from time limits; jobs that end early or late make the estimate wrong until the next recheck.
• No memory request → 512 MB assumed; such jobs are not pinned.
• Time limits taken at face value (time_aware).
• Over-cap jobs run to completion (QOS limits and flex).
• One sjpd per cluster; on a backup controller sjp.place() returns false.
• Upgrades need scontrol reconfigure.

Known bugs and feature requests: Issues.