storage systems, compilers, … all will be useful. A talk I gave at Stanford a while back has some material on this (although focuses on building systems in the context of more traditional IR systems rather than ML systems).
Building Systems for ML: Storage, Compilers, and IR Design
By
–