This post traces the decline of the Linux page cache in modern databases and storage systems. It explores the shift from mmap() and buffered I/O to O_DIRECT, driven by SSD endurance improvements, NVMe Zoned Namespaces, and user-space filesystems. Featuring case studies of MySQL, ScyllaDB, and Redpanda, it highlights why these systems bypass the kernel cache to achieve predictable p99 latencies and tighter control over I/O.

By Codemia • Aug 20, 2025

8 min read

This article explores how Non-Uniform Memory Access (NUMA) is reshaping microservice placement. It traces the evolution from early SMP kernels to modern Linux features like cgroups v2 and eBPF locality hints, and highlights Kubernetes topology-aware scheduling. Real-world data shows that aligning pods and stateful services with NUMA nodes can significantly reduce tail latency and improve throughput, making NUMA awareness as critical as network awareness in distributed system design.

By Codemia • Aug 17, 2025

5 min read

This article traces the evolution of Linux asynchronous I/O from epoll to io_uring, highlighting how multishot receive operations streamline network event handling. It explains why multishot removes much of the boilerplate and syscall overhead of traditional event loops, and presents benchmarks showing io_uring’s throughput gains and tail latency reductions over epoll—making 2025 the year to rethink event-driven server design.

By Codemia • Aug 14, 2025

5 min read

Linux CPU Scheduling Evolution and Facebook's BOLT: Cutting Latency with sched_ext

By Codemia • Aug 13, 2025

5 min read