Print seed diagnostics in random-seed tests

Author: SaridakisStamatisChristos

kll_sketch/tests/test_random_seeds.py

@@ -0,0 +1,74 @@
+"""Regression tests that exercise the sketch under multiple RNG seeds."""
+
+from __future__ import annotations
+
+import math
+import random
+import statistics
+from typing import Iterable
+
+import pytest
+
+from kll_sketch import KLL
+
+
+def _exact_quantile(values: Iterable[float], q: float) -> float:
+    ordered = sorted(values)
+    if not ordered:
+        raise ValueError("_exact_quantile requires a non-empty iterable")
+    # Linear interpolation between neighbouring order statistics to mirror the
+    # behaviour of NumPy's ``quantile`` default method.
+    position = q * (len(ordered) - 1)
+    lower = math.floor(position)
+    upper = math.ceil(position)
+    if lower == upper:
+        return ordered[lower]
+    weight_upper = position - lower
+    weight_lower = 1.0 - weight_upper
+    return ordered[lower] * weight_lower + ordered[upper] * weight_upper
+
+
+@pytest.mark.parametrize("seed", [3, 17, 221, 1987, 4096])
+def test_quantile_accuracy_across_random_seeds(seed: int) -> None:
+    rng = random.Random(seed)
+    samples = [rng.gauss(0.0, 1.0) for _ in range(8_000)]
+
+    sketch = KLL(capacity=256, rng_seed=seed)
+    sketch.extend(samples)
+
+    approx = sketch.quantile(0.5)
+    exact = _exact_quantile(samples, 0.5)
+    dispersion = statistics.pstdev(samples)
+    tolerance = max(1e-9, 0.12 * dispersion)
+
+    print(
+        "seed={seed}: approx median={approx:.6f}, exact={exact:.6f}, tolerance={tolerance:.6f}".format(
+            seed=seed, approx=approx, exact=exact, tolerance=tolerance
+        )
+    )
+
+    assert abs(approx - exact) <= tolerance
+
+
+def test_deterministic_compactions_for_fixed_seed() -> None:
+    seed = 123_456
+    rng = random.Random(seed)
+    payload = [rng.uniform(-5.0, 5.0) for _ in range(5_000)]
+
+    a = KLL(capacity=200, rng_seed=seed)
+    b = KLL(capacity=200, rng_seed=seed)
+    for value in payload:
+        a.add(value)
+        b.add(value)
+
+    print(
+        "deterministic compaction: sketch bytes size={size}, levels={levels}".format(
+            size=len(a.to_bytes()),
+            levels=sum(len(level) for level in a._levels),
+        )
+    )
+
+    assert a._levels == b._levels
+    assert a.to_bytes() == b.to_bytes()
+    for q in [0.05, 0.5, 0.95]:
+        assert math.isclose(a.quantile(q), b.quantile(q), rel_tol=1e-12, abs_tol=1e-12)