The problem
Can a custom scheduler be tested without turning the experiment into an unbounded production risk?
What current research shows
Linux sched_ext lets BPF programs define scheduler behavior, can be enabled or disabled dynamically, and restores default scheduling when errors or stalls are detected. The official documentation also states that the BPF scheduler APIs have no stability guarantees between kernel versions.
Where the evidence stops
A sched_ext experiment isn’t a stable kernel ABI, a universal performance result, or proof that a custom scheduler will survive version changes.
What Valen Systems is testing
Version every scheduler, record whether sched_ext was active, verify fallback behavior, and compare equal work under the same host and kernel conditions.