mirror of
https://github.com/hanzoai/benchmarks.git
synced 2026-10-30 14:59:11 +00:00
zeekay
5afa0b2829
Separate, orthogonal axis from the zaphttp framing-CPU test (that carries the body opaque and never touches JSON). Isolates ZAP's real structural win: typed binary fields read zero-copy (no scan/reflect/per-field alloc) vs HTTP+JSON (de)serializing text every call. One eth-style record, same compute, same field set both sides; JSON held to its fastest libs (goccy, sonic) + stdlib. New shared `record` package (DRY: one record + codecs used by the microbench, the /rpc server, and the ZAP load generator). serde_test.go: decode/encode/ round-trip/handler benches + wire sizes + round-trip equivalence. main.go: POST /rpc route (zap transport = zero-copy decode; http transport = JSON codec). zapload: -scenario rpc sends the identical ZAP-typed body. Numbers (GB10 arm64, go1.26, median of 3, default GOGC/GOMAXPROCS): - Decode: ZAP 13.4ns/0-alloc vs goccy 473ns/4, sonic 845ns/8, stdlib 2334ns/13 (35–174x, 0 vs many allocs) — the crush. - Encode: honest wash — ZAP 629ns/12-alloc vs goccy 424ns/4 (v1.3.0 SetBytes defers tail via copy + offsets slice; SetBytesFixed is luxfi/zap-only). - Round trip: ZAP 648ns vs goccy 904 / sonic 1509 / stdlib 3107. - Handler (per-req): ZAP 271ns/5-alloc vs goccy 431/4, sonic 736/7, stdlib 1444/9. - Wire: total ZAP 162B vs JSON 233B. - End-to-end /rpc c125 d6s: ZAP 534,945 rps vs sonic 483k (1.11x), goccy 452k (1.18x), stdlib 407k (1.31x). Reproducible off published tags (zap-proto/go v1.3.0, no local replaces).