@effect-bench/node 0.0.0 → 0.0.1

package/README.md

@@ -0,0 +1,88 @@
+# effect-bench
+
+Benchmarks for [Effect](https://effect.website) code, with JSON output and baseline verdicts built for agents and CI.
+
+Community project, not affiliated with the Effect team.
+
+```ts
+// bench/schema.bench.ts
+import { Bench } from '@effect-bench/core'
+import { Effect } from 'effect'
+
+export default Bench.suite('workflowCompiler', { layer: NodeRegistry.layer })(
+  Bench.params('compile', { nodes: [10, 100, 1000] }, {
+    prepare: ({ nodes }) => Effect.succeed(makeDraft(nodes)),   // untimed
+    run: draft => compile(draft),                                // timed
+  }),
+  Bench.group('decode', { baseline: 'sync' })(
+    Bench.sync('sync', () => decodeSync(lead)),
+    Bench.make('effect', () => decodeEffect(lead)),
+  ),
+  Bench.sync('encode', () => encode(payload), { budget: { p99: '5 millis' } }),
+)
+```
+
+```bash
+effect-bench list
+effect-bench run --save before          # store a baseline
+effect-bench run --compare before --json
+effect-bench check --compare before     # exit 1 on a regression or blown budget
+effect-bench run --rigorous --forks 8   # more processes for a tighter interval
+```
+
+## Writing benches
+
+| Constructor | Measures |
+| --- | --- |
+| `Bench.make(name, effect)` | re-runs one Effect value |
+| `Bench.make(name, ctx => effect)` | builds and runs a fresh Effect per operation |
+| `Bench.sync(name, ctx => value)` | a plain function |
+| `Bench.params(name, matrix, { prepare?, run \| sync })` | one bench per combination, ids like `compile[nodes=100]` |
+| `Bench.group(name, { baseline })(...)` | alternatives, reported as a ratio to the baseline |
+
+`Bench.suite(name, { layer, setup, ...options })` builds `layer` once per suite and runs `setup` in the suite's scope; its result is the `ctx` each bench receives. Benches can only use services the layer provides — anything else is a type error.
+
+Options (suite or bench): `samples`, `minTime`, `warmup`, `memory` (GC + heap delta per sample), `budget: { p50, p99 }`, `tags`, and `check` (validates one result before timing).
+
+## Output
+
+`--json` prints one record per line on stdout: a `result` per bench and a final `summary`. Logs go to stderr.
+
+```json
+{"type":"result","schemaVersion":1,"status":"ok","id":"leadSchema/decode/effect","stats":{"medianNs":858,"p99Ns":943,"opsPerSec":1165161,...},"comparison":{"baseline":"before","change":0.051,"ci":[0.039,0.058],"verdict":"slower","stale":false},"relative":{"baselineId":"leadSchema/decode/sync","ratio":5.1}}
+{"type":"summary","schemaVersion":1,"total":4,"failed":0,"regressions":1,"budgetsExceeded":0,"stale":0,"exitCode":1}
+```
+
+## How verdicts work
+
+The CLI measures each suite in several fresh processes (`--forks`, default 3; 6 with `--rigorous`), interleaved fork by fork so slow drift spreads across suites. Drift between processes — JIT decisions, CPU frequency — is much larger than noise within one run, so each fork's median counts as one observation: the 95% interval for the change is a pooled t-interval on log medians.
+
+`faster`/`slower` only when the whole interval is beyond the noise threshold (`--threshold`, default 3%), `same` when it is inside, `inconclusive` otherwise. On a laptop with 3 forks per side, repeated comparisons of unchanged code gave no false verdicts in 96 tries; a 15% slowdown was flagged every time and a 5% one most of the time (otherwise `inconclusive`). Treat `inconclusive` as "measure more": `--rigorous` or more `--forks`.
+
+`--forks 0` measures in the CLI process, and `Runner.run` does the same when called from code. With a single fork on both sides the interval only sees within-run noise and will overstate confidence.
+
+Exit codes: `0` ok, `1` regression or blown budget, `2` failed bench or error, `3` baseline recorded on another runtime or machine.
+
+## Config
+
+Optional `effect-bench.config.ts` in the working directory:
+
+```ts
+import { BenchConfig } from '@effect-bench/cli'
+
+export default BenchConfig.make({
+  include: ['**/bench/*.bench.ts'],
+  modes: { ci: { samples: 50, minTime: '1 second', forks: 5 } },
+  noiseThreshold: 0.05,
+})
+```
+
+## Packages
+
+- `@effect-bench/core` — DSL, measuring loop, stats, comparison, runner, reporters. No platform imports.
+- `@effect-bench/node` — hrtime clock, GC and heap, environment fingerprint, file baseline store.
+- `@effect-bench/cli` — `effect-bench` command.
+
+## Known limitations
+
+- Bench files are loaded with Node's built-in type stripping, so they must use erasable TypeScript syntax only.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@effect-bench/node",
-  "version": "0.0.0",
+  "version": "0.0.1",
   "type": "module",
   "exports": {
     ".": {
@@ -9,7 +9,7 @@
     }
   },
   "dependencies": {
-    "@effect-bench/core": "0.0.0"
+    "@effect-bench/core": "0.0.1"
   },
   "peerDependencies": {
     "effect": "^4.1.2",
@@ -25,5 +25,6 @@
   ],
   "publishConfig": {
     "access": "public"
-  }
+  },
+  "scripts": {}
 }