microlinkhq / keyvhq / 40593712452

97%

push

github

perf(core): drop json-buffer, serialize with native JSON (#283)

* perf(core): serialize with JSON.stringify replacer instead of json-buffer

json-buffer rebuilds every object by hand in JS (string concatenation
plus a JSON.stringify call per key and per leaf). A replacer passed to
the native JSON.stringify keeps the exact same encoding: Buffers become
":base64:<data>" (checked on the raw holder value, before Buffer#toJSON
runs) and strings starting with ":" get one more ":". Parsing keeps the
same reviver logic without the regexes. json-buffer is dropped from
@keyvhq/core dependencies.

A/B, node benchmark/index.js, 5 alternating runs per side, median of
medians (node 24.16.0, arm64, load avg 7-9):

| case | base | head | delta |
|---|---|---|---|
| core set+get object | 311.5 ms | 198.1 ms | -36.4% |
| core get buffer | 141.4 ms | 140.1 ms | -0.9% |
| core namespace set+get object | 311.7 ms | 228.3 ms | -26.8% |
| memoize hit | 236.0 ms | 231.0 ms | -2.1% |
| memoize concurrent dedupe | 51.0 ms | 51.7 ms | +1.4% |
| memoize stale | 111.0 ms | 90.7 ms | -18.3% |

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* perf(core): revive parsed values with a walk instead of a JSON.parse reviver

Passing any reviver to JSON.parse makes V8 leave its fast path: a no-op
reviver alone costs 5-6x on the benchmark payloads (158 -> 877 ns for
the memoize payload, 461 -> 2936 ns for the object payload). Parsing
plainly and then walking the result to unescape ":" strings and decode
":base64:" Buffers does the same work in 208 and 722 ns. Keys come from
Object.keys, so enumerable properties on Object.prototype are not
visited, and a value is only written back when it changed.

Adds a test that pins the stored bytes to the json-buffer format.

A/B against master, node benchmark/index.js, 5 alternating runs per
side, median of medians (node 24.16.0, arm64, load avg 15-35):

| case | base | head | delta |
|---|---|---|---|
| core set+get object | 300.3 ms | 84.... (continued)

342 of 368 branches covered (92.93%)

Branch coverage included in aggregate %.

86 of 86 new or added lines in 1 file covered. (100.0%)

1015 of 1035 relevant lines covered (98.07%)

64.74 hits per line