After a Linux kernel upgrade caused latency regressions in its ads-serving infrastructure, Meta deployed sched_ext, an open-source BPF-based kernel scheduler, to write a custom scheduling policy that soft-partitions CPUs between latency-critical and non-critical workloads. The custom policy delivered a 28% reduction in p99 service latency, 3.28 megawatts of power savings, and a 1.1% increase in ads ranked, and a subsequent user-space policy update produced a further 60% latency reduction and 18% fewer timeout errors. The post frames this as a case for workload-specific kernel scheduling that bypasses the multi-year upstream Linux kernel release cycle.
