feat: add soak runner with budget guards and invariant scoreboard

2026-07-21 01:23:40 -04:00
parent cca7071dbd
commit 9aa97a61ae
3 changed files with 416 additions and 0 deletions
@@ -129,3 +129,66 @@ class TestWeightedMix:
    def test_invalid_weights_rejected(self):
        with pytest.raises(ValueError):
            weighted_mix({"P1": 0.0}, lambda: 0.5)
+
+
+# ═══════════ T11: 记分板硬不变量(纯函数) ═══════════
+
+from tools.soak.scoreboard import (  # noqa: E402
+    inv_call_ids_unique,
+    inv_rows_match_calls,
+    inv_rpm_never_exceeded,
+    inv_rss_stable,
+    structured_success_rate,
+)
+
+
+def _row(call_id, *, source="s1", created_at="2026-07-20T10:00:00", error=None, session="r-p3"):
+    return {
+        "call_id": call_id,
+        "source_name": source,
+        "created_at": created_at,
+        "error": error,
+        "session_id": session,
+    }
+
+
+class TestInvariants:
+    def test_rows_match_calls(self):
+        rows = [_row("a"), _row("b")]
+        inv_rows_match_calls(rows, expected_calls=2)
+        with pytest.raises(AssertionError):
+            inv_rows_match_calls(rows, expected_calls=3)
+
+    def test_call_ids_unique(self):
+        inv_call_ids_unique([_row("a"), _row("b")])
+        with pytest.raises(AssertionError):
+            inv_call_ids_unique([_row("a"), _row("a")])
+
+    def test_rpm_minute_bucket(self):
+        # 固定分钟窗口口径(与限流器 int(sec/60) 同源;滑动窗会对合法的
+        # 跨窗背靠背流量误报,不采用)
+        rows = [
+            _row("a", created_at="2026-07-20T10:00:01"),
+            _row("b", created_at="2026-07-20T10:00:59"),
+            _row("c", created_at="2026-07-20T10:01:01"),
+        ]
+        inv_rpm_never_exceeded(rows, {"s1": 2})
+        rows.append(_row("d", created_at="2026-07-20T10:00:30"))
+        with pytest.raises(AssertionError):
+            inv_rpm_never_exceeded(rows, {"s1": 2})
+
+    def test_rpm_ignores_unlimited_sources(self):
+        inv_rpm_never_exceeded([_row(str(i)) for i in range(100)], {"s1": 0})
+
+    def test_rss_stable(self):
+        inv_rss_stable([100.0, 105.0, 110.0], max_growth_mb=50.0)
+        with pytest.raises(AssertionError):
+            inv_rss_stable([100.0, 400.0], max_growth_mb=50.0)
+
+    def test_structured_success_rate(self):
+        rows = [
+            _row("a"),
+            _row("b", error="ResultInvalidError: x"),
+            _row("c", session="r-p1"),  # 非 P3 剔除
+        ]
+        assert structured_success_rate(rows, session_suffix="-p3") == pytest.approx(0.5)