Turso exposes io_uring’s hidden readahead tax

The lesson is less “io_uring is faster” than “async I/O needs an explicit queueing strategy”: bypassing the page cache trades kernel convenience for application responsibility.

• –A 32-page prefetch window enabled 91–93% request merging, turning tiny reads into larger device requests.
• –O_DIRECT avoids copying through the page cache but can shift the cost into CPU cache misses.
• –SQPOLL helps when spare vCPUs exist; otherwise, its polling thread competes with application work.
• –Turso’s plain io_uring path can batch submissions, but its pager still waits one page at a time.
• –Database and AI-serving workloads should benchmark latency, CPU, cache behavior, and device merges—not just IOPS.