Consumable system features would open up a lot of possibilities for better load balancing, especially for hydra build slaves. Examples of a consumable feature could be a CPU or a gigabyte of memory. Currently there is not really any good way to do load balancing for derivations for anything other than CPU usage, and consumable features could be a really great way to allow users to do this effectively.

It could be specified with a feature:num syntax, with the lack of the :num or :0 indicating an infinite (non-consumable) feature (i.e., what all features are right now).

Example build machines file with scratchy having 32 "mem" features (e.g., it might have 32GB of memory):

[email protected]  i686-linux      /home/nix/.ssh/id_scratchy_auto        8 1 kvm,mem:32
[email protected]     i686-linux      /home/nix/.ssh/id_scratchy_auto        8 2
[email protected]   i686-linux      /home/nix/.ssh/id_scratchy_auto        1 2 kvm benchmark

A derivation should likewise be able to specify how many of each consumable feature it consumes. e.g., could use this syntax:

requiredSystemFeatures = [ "kvm" "mem:4" ];

or maybe this would be better:

requiredSystemFeatures = [ "kvm" "mem" "mem" "mem" "mem" ];

This can be useful for managing any kind of specialized hardware resource exposed / accessible in the sandbox, e.g., if a sandbox-mapped device has a limited number of handles that can be opened at a time.