PREEMPT_NONE Is Dead; Your Postgres Probably Doesn’t Care
Curated from Lobsters
For years, the PostgreSQL community relied on the kernel’s PREEMPT_NONE configuration to guarantee low-latency query execution, assuming that preventing preemption was the primary lever for performance stability. This article dismantles that long-held assumption by demonstrating that modern Linux kernels, even with standard preemptible configurations, do not inherently degrade database throughput in ways that justify the operational complexity of building custom, low-preemption kernels. The focus has shifted from fighting the scheduler to optimizing the database engine itself, where configuration tuning and workload design yield far more reliable results than kernel-level hacks. This represents a significant paradigm shift for SREs who have invested heavily in specialized infrastructure to mitigate latency spikes. Relying on outdated kernel myths often leads to fragile, hard-to-maintain environments that offer diminishing returns. Stop optimizing the scheduler and start optimizing your queries.
p a href="https://lobste.
— Lobsters