Zig

Zig is in this suite because comptime reflection (@typeInfo) is a different implementation model from Java/Kotlin runtime reflection, C# source generation, or Rust derive macros. The runner times official std.json (typed parseFromSlice versus streaming Scanner / parseFromTokenSource) against an in-tree comptime byte-packed baseline, against serde.zig, a format-agnostic framework that uses the same @typeInfo walk for JSON, MessagePack, YAML, TOML, and ZON, and against schema codecs (Protocol Buffers, FlatBuffers, Cap’n Proto) generated from the shared suite IDLs.

Runtime

What it is

Zig compiles to native machine code. There is no hidden virtual machine and no garbage collector. Allocation is explicit: the benchmark runner passes allocators in. Comptime means the compiler can run Zig code while it builds. @typeInfo walks a struct at compile time instead of using Java-style reflection at run time.

	This suite
Compiler	Zig 0.16.x. Version 0.15 or 0.17 will not build this tree.
Prepare	`./scripts/install-host-requirements.sh zig` installs into `~/.local/zig`
Run	`zig/scripts/run-benchmarks.sh` (`zig build`)
Memory	Explicit allocators. No garbage collector.

What this suite runs

Zig still changes in breaking ways between minor versions, so the host script installs 0.16.x and the checker rejects any other series. The Zig binary does not start Python. Cap’n Proto uses the official C++ library, because Zig 0.16 has no native Cap’n Proto plugin.

What changes the numbers

Building without optimizations is the error that changes the numbers the most, because an unoptimized Zig binary is far slower than a release build. Codecs that use comptime, such as std.json, serde.zig, and comptime-bin, generate the field walk while the program compiles.

A @bitCast of a live suite value is not a valid encoding. Slices inside that value are pointers, not payload bytes. See Not a @bitCast.

Suite-specific gotchas

The capnproto row needs libcapnp and libkj under ~/.local.

Some serde.zig rows support only the message and strings data types.

These times cannot be ranked against another language.

Where to go next

The steps to install the toolchain and run the benchmark are in zig/README.md. The language overview is Zig overview.

Benchmark runner

Directory: zig/ (repository root)
Output: logs/zig/YYYY-MM-DD-HHMMSS.csv (Language=zig, times in nanoseconds)
Runner: zig/scripts/run-benchmarks.sh {smoke|all-single|full|research}
Registration: zig/src/serializers.zig

The shell script resolves the run config to JSON. The Zig binary does not spawn Python. Prepare is untimed. The harness owns a reusable output buffer per serializer. Timed I/O is serialize plus deserialize only. Schedule is SHA-256 + SplitMix64 Fisher–Yates (golden vector C, B, A).

Serializers (wired)

Name	Category	Package	Stream	Notes
std.json	JSON	std	text_on_stream	`Stringify.value` + `parseFromSlice` into the suite struct
std.json.scanner	JSON	std	text_on_stream	Same stringify; decode is `Scanner` + `parseFromTokenSource`
std.zon	ZON	std	text_on_stream	Official `std.zon.stringify` + `std.zon.parse.fromSliceAlloc`
comptime-bin	Binary	in-tree	adapted	Comptime field walk; LE ints; `u32` length + bytes for strings
serde.json	JSON	serde.zig 1.0.7	adapted	One comptime API, JSON path
serde.msgpack	Binary	serde.zig 1.0.7	adapted	Same API, MessagePack
serde.yaml	Text	serde.zig 1.0.7	text_on_stream	Same API, YAML. message / strings only
serde.toml	Text	serde.zig 1.0.7	text_on_stream	Same API, TOML
serde.zon	Text	serde.zig 1.0.7	text_on_stream	Same API, Zig Object Notation
serde.xml	Text	serde.zig 1.0.7	text_on_stream	Same API, XML. message / strings only
zig-msgpack	Binary	zigcc/zig-msgpack 0.0.14	adapted	Official MessagePack Payload API
msgpack.zig	Binary	lalinsky/msgpack.zig 0.7.0	native	Typed `encode` / `decodeFromSlice`
zbor	Binary	r4gus/zbor 0.21.0	adapted	Native Zig CBOR (`stringify` / `parse`)
s2s	Binary	ziglibs/s2s	native	Native binary “struct to stream”
protobuf	Schema	Arwalk/zig-protobuf 5.0.0	adapted	Generated from `schemas/v2/protobuf/benchmark_v2.proto`. Prepare copies suite → generated message; timed path is `encode` / `decode`
flatbuffers	Schema	nDimensional/zig-flatbuffers 0.2.1	adapted	Generated from `cpp/schemas/benchmark.fbs`. Timed path is `Builder.writeTable` / `decodeRoot`
capnproto	Schema	Cap’n Proto C++ 1.0.2	adapted	Generated from `cpp/schemas/benchmark.capnp`. Official C++ runtime via a C ABI (same pattern as Swift). Zig 0.16 has no native plugin

Not a `@bitCast` of the whole fixture

A live suite value has []const u8 slices. @bitCast of that type is not a portable encoding. comptime-bin is the honest idiomatic stand-in: comptime reflection writes a length-prefixed little-endian image.

Schema generation

These rows compile the same suite IDLs as the other languages. They do not invent Zig-only schemas.

Row	Shared schema	Regenerate
protobuf	`schemas/v2/protobuf/benchmark_v2.proto`	`./zig/scripts/generate-protobuf.sh`
flatbuffers	`cpp/schemas/benchmark.fbs`	`./zig/scripts/generate-flatbuffers.sh`
capnproto	`cpp/schemas/benchmark.capnp`	`./zig/scripts/generate-capnp.sh`

Prepare copies each suite fixture into the library’s native form (untimed). The timer measures encode and decode only. Fidelity copies back to the suite struct after decode.

capnproto uses the official C++ library (like Swift) because the Zig Cap’n Proto plugin requires Zig 0.17-dev. The host needs libcapnp / libkj under ~/.local (./scripts/install-host-requirements.sh zig).

Dashboard

Zig