json-experiment

An experimental, performance-focused JSON marshaler for Go.

Benchmarks

Benchmark 5 compares:

• MarshalAppend, which reuses caller-provided destination capacity.
• Marshal, which returns an independently owned byte slice.
• Go's encoding/json.
• Go's experimental encoding/json/v2 using MarshalWrite.
• sonic.ConfigFastest.Marshal.
• Sonic's reusable-buffer EncodeInto API.

For like-for-like comparisons, MarshalAppend corresponds to MarshalWrite and EncodeInto, while Marshal corresponds to the marshal APIs that return owned output. The append path can avoid output allocation when the destination has sufficient capacity.

Benchmark 5

This is a five-run median on Go 1.26 using GOAMD64=v3 with the JSON v2 and SIMD experiments enabled. The suite includes maps, primitive slices, numbers, struct variants, UTF-8 validation, and standard marshaling interfaces.

GOAMD64=v3 GOEXPERIMENT=jsonv2,simd go test \
  -benchmem -run='^$' -count=5 -bench='^Benchmark' .

Five-run median latency (lower is better):

Workload	MarshalAppend	Marshal	encoding/json	json/v2 Write	Sonic Marshal	Sonic EncodeInto
map[string]int	192.8 ns	440.3 ns	1232 ns	734.8 ns	661.4 ns	424.1 ns
map[string][]int	210.9 ns	502.4 ns	1571 ns	888.4 ns	718.4 ns	433.2 ns
map[string]any	320.8 ns	567.9 ns	2566 ns	1848 ns	940.6 ns	544.2 ns
[]int	72.52 ns	204.6 ns	404.6 ns	293.7 ns	358.9 ns	186.0 ns
float32	36.80 ns	65.93 ns	156.7 ns	142.6 ns	130.7 ns	93.96 ns
float64	56.63 ns	108.9 ns	204.5 ns	171.9 ns	179.4 ns	104.7 ns
mixed struct	274.5 ns	627.4 ns	1622 ns	1075 ns	888.5 ns	613.1 ns
struct slice with metadata maps	810.1 ns	1536 ns	4167 ns	2898 ns	2374 ns	1484 ns
quoted struct fields	190.4 ns	344.4 ns	1411 ns	671.3 ns	511.1 ns	273.3 ns
omitempty / omitzero	97.64 ns	259.1 ns	859.0 ns	776.4 ns	422.3 ns	261.3 ns
UTF-8 validation: ASCII	120.4 ns	809.7 ns	2006 ns	620.5 ns	1099 ns	166.8 ns
UTF-8 validation: Unicode	236.5 ns	921.1 ns	1325 ns	579.8 ns	951.3 ns	170.1 ns
UTF-8 validation: invalid byte	272.5 ns	1564 ns	2227 ns	740.5 ns	2867 ns	3501 ns
encoding.TextMarshaler	44.08 ns	81.16 ns	190.5 ns	142.4 ns	139.2 ns	105.5 ns
json.Marshaler	33.30 ns	66.64 ns	137.7 ns	122.3 ns	120.2 ns	110.5 ns

Median allocations per operation:

Workload	MarshalAppend	Marshal	encoding/json	json/v2 Write	Sonic Marshal	Sonic EncodeInto
map[string]int	0	1	11	3	3	2
map[string][]int	0	1	10	3	3	2
map[string]any	0	1	23	14	3	2
[]int	0	1	3	2	3	2
float32	0	1	3	2	3	2
float64	0	1	3	2	3	2
mixed struct	0	1	7	4	4	3
struct slice with metadata maps	0	1	19	10	7	6
quoted struct fields	0	1	6	2	3	2
omitempty / omitzero	0	1	3	2	3	2
UTF-8 validation: ASCII	0	1	3	2	3	2
UTF-8 validation: Unicode	0	1	3	2	3	2
UTF-8 validation: invalid byte	0	1	3	2	3	4
encoding.TextMarshaler	1	2	2	1	3	2
json.Marshaler	1	2	2	1	3	2

The custom MarshalerAppend precedence benchmark only applies to this package. Its median results were 16.74 ns with zero allocations for MarshalAppend, and 47.32 ns with one allocation for Marshal.

The complete Benchmark 5 output, including bytes and allocations per operation, is available in bench5.txt. Benchmarks 1–4 and their original commentary are preserved in README-old.md.

---

Results vary by hardware, Go version, and workload. Run the benchmarks on the target system before drawing conclusions for a particular application.