Longer term, the major selling proposition of on-demand would be low latency. With the current implementation, we process inherents first and then extrinsics. This means at the time of scheduling we are not yet aware of any orders contained in the current block and thus scheduling of them will only happen on the next block. We therefore have at least 6s latency for processing an on-demand order, even if cores are empty.

The actual scheduling happens during processing of the paras inherent. We would need to move the scheduling to after we have processed extrinsics, so we can take fresh orders into account immediately.