handle large magnitude negative floats consistently

src/whylogs/core/summaryconverters.py

@@ -124,6 +124,9 @@ def single_quantile_from_sketch(sketch: kll_floats_sketch, quantile: float):
 def _calculate_bins(end: float, start: float, n: int, avg_per_bucket: float, max_buckets: int):
     # Include the max value in the right-most bin
     end += abs(end) * 1e-7
+    abs_end = abs(end)
+    abs_start = abs(start)
+    max_magnitude = max(abs_end, abs_start)
 
     # the kll_floats_sketch use 32bit floats, so we check precision against np.float32
     float_mantissa_bits = np.finfo(np.float32).nmant
@@ -133,7 +136,7 @@ def _calculate_bins(end: float, start: float, n: int, avg_per_bucket: float, max
     width = (end - start) / n_buckets
 
     # check for precision of width with respect to float_mantissa_bits and bin width:
-    bits_in_max = math.floor(math.log2(end))
+    bits_in_max = math.floor(math.log2(max_magnitude))
     width_bits = math.floor(math.log2((end - start) / n_buckets))
     logger.debug(f"bits_in_max is: {bits_in_max}")
     logger.debug(f"width_bits is: {width_bits}")
@@ -143,7 +146,7 @@ def _calculate_bins(end: float, start: float, n: int, avg_per_bucket: float, max
         logger.info(f"Width must be larger than {bits_in_width} bits.")
         new_buckets = math.floor((end - start) / math.pow(2, bits_in_width))
         logger.warn(f"Avoiding bin edge collisions by resizing to {new_buckets} buckets")
-        n_buckets = new_buckets
+        n_buckets = max(new_buckets, 1)
         width = (end - start) / n_buckets
 
     # Calculate histograms from the Probability Mass Function

tests/unit/app/test_session.py

@@ -1,5 +1,6 @@
 from logging import getLogger
 
+import pandas as pd
 import pytest
 
 from whylogs.app.config import SessionConfig
@@ -78,6 +79,19 @@ def test_session_profile_small(df_single):
         assert len(flat_summary) == 1
 
 
+def test_session_profile_negative_ints():
+    df = pd.DataFrame(range(-22335544310, -22335542310), columns=["negative"])
+    session = session_from_config(SessionConfig("default-project", "default-pipeline", [], False))
+    profile = session.log_dataframe(df)
+    TEST_LOGGER.debug(f"logged df: {df.shape} ")
+    TEST_LOGGER.debug(f"logged profile: {profile} ")
+    summary = profile.flat_summary()
+    TEST_LOGGER.debug(f"logged summary: {summary} ")
+    flat_summary = summary["summary"]
+    TEST_LOGGER.debug(f"logged flat_summary: {flat_summary} ")
+    assert len(flat_summary) == 1
+
+
 def test_session_profile_two_column(df_two_int_col):
     TEST_LOGGER.debug(f"About to log {df_two_int_col.describe()} with columns {df_two_int_col.columns}")
     session = session_from_config(SessionConfig("default-project", "default-pipeline", [], False))