Eight Bytes Are a Number

A practical guide to zero-copy parsing in Rust shows how treating bytes in-place as already-formed numbers avoids allocations and vastly improves throughput for binary protocols. A command enum (EngineCommand) is designed so every variant is composed of Copy integer types and has a fixed-size layout: for example NewLimit is 57 bytes, CancelByOrder 16 bytes. The decoder checks the envelope kind and length, then reads fields from known offsets (read_u64, read_i64, read_u128, decode_side, etc.) to construct variants without owning buffers. The same fixed-offset approach is used for encoding, writing directly into a caller-owned payload region so neither parsing nor serialization requires intermediate heap allocations.

Concrete measurements back the claim: per-command decode times around 4.1 ns and throughput near 237-245 Melem/s. A verification using two small Rust programs and strace contrasts parse_id (which copies bytes into a Vec and reserves memory) with zero_copy_parse_id (which uses from_le_bytes on the slice). The copying version triggers extra brk/mmap/munmap activity (including a 1,052,672-byte mmap for the reserved Vec) while the zero-copy variant avoids that work, demonstrating that “eight bytes are already a number” and that preserving non-owning, Copy types yields dramatic syscall and allocation savings.