vector_of_kll_floats_sketches.get_quantiles() returns wrong values with float32

[tyler-rt]

#!/usr/bin/env python3
"""
Minimal example: vector_of_kll_floats_sketches.get_quantiles() returns WRONG VALUES with float32
"""
import numpy as np
from datasketches import vector_of_kll_floats_sketches

# Create test data: 1000 samples between -100 and -10
np.random.seed(42)
test_data = np.random.uniform(-100, -10, size=(1000, 1)).astype(np.float32)

print("Test data: 1000 samples between -100 and -10")
print(f"True min: {test_data.min():.2f}, True max: {test_data.max():.2f}")

# Create sketch and add data
kll = vector_of_kll_floats_sketches(200, 1)
kll.update(test_data)

# Request p0.0001 (should be ~-100) and p0.9999 (should be ~-10)
ranks_list = [0.0001, 0.9999]
ranks_array32 = np.array(ranks_list, dtype=np.float32)
ranks_array64 = np.array(ranks_list, dtype=np.float64)


print("\n" + "="*60)
print("BUG: numpy array with dtype=np.float32 returns WRONG quantiles")
print("="*60)

quants_array = kll.get_quantiles(ranks_array32)
print(f"\nWith numpy array with dtype=np.float32: {ranks_array32}")
print(f" p0.0001 = {quants_array[0][0]:.2f} (expected: ~-100)")
print(f" p0.9999 = {quants_array[0][1]:.2f} (expected: ~-10)")
print(f" ✗ WRONG: Both values near minimum!")

quants_array64 = kll.get_quantiles(ranks_array64)
print(f"\nWith numpy array with dtype=np.float64: {ranks_array64}")
print(f" p0.0001 = {quants_array64[0][0]:.2f} (expected: ~-100)")
print(f" p0.9999 = {quants_array64[0][1]:.2f} (expected: ~-10)")
print(f" ✓ CORRECT")

Test data: 1000 samples between -100 and -10
True min: -99.58, True max: -10.03

============================================================
BUG: numpy array with dtype=np.float32 returns WRONG quantiles
============================================================

With numpy array with dtype=np.float32: [1.000e-04 9.999e-01]
 p0.0001 = -98.69 (expected: ~-100)
 p0.9999 = -99.50 (expected: ~-10)
 ✗ WRONG: Both values near minimum!

With numpy array with dtype=np.float64: [1.000e-04 9.999e-01]
 p0.0001 = -99.50 (expected: ~-100)
 p0.9999 = -10.28 (expected: ~-10)
 ✓ CORRECT

[jmalkin]

My employer (and long may they keep me!) has pretty strict limits on interaction with this project at the moment so I'm currently unable to contribute a fix or anything. But I believe the issue is that even though the sketch is set up in terms of float (np.float32) the definition of get_quantiles() uses double and it's not casting the array properly.

[c-dickens]

Thanks for flagging @tyler-rt. I'm currently in the process of moving house so will try to take a look at this when I get some free time. Will need a degree of short term patience to give it proper attention though.