A search service (Marginalia Search) gutted its old index internals and dropped memory-mapped B-trees. In their place: a deterministic, block-aligned skip list tuned for direct reads on NVMe SSDs.
It runs on 128KB block sizes, uses custom buffer pools, and leans hard on io_uring for async position lookups. The payoff? Noticeably faster reads and cleaner latency across the board.
Why it matters: More systems are ditching mmap and building SSD-first, hardware-aware data structures. Marginalia just joined the modern camp.
