Fixed flaky test

bench/gen_bench_expectations.py

Index: bench/gen_bench_expectations.py
diff --git a/bench/gen_bench_expectations.py b/bench/gen_bench_expectations.py
index 5a9e1e495b86c4d97d6151fdeaeb205cdbab07c7..af0e46509163136e5514b70e5b0091ba0235acc3 100644
--- a/bench/gen_bench_expectations.py
+++ b/bench/gen_bench_expectations.py
@@ -46,13 +46,14 @@ def compute_ranges(benches):
   Returns:
     a list of float [lower_bound, upper_bound].
   """
-  minimum = min(benches)
-  maximum = max(benches)
-  diff = maximum - minimum
   avg = sum(benches) / len(benches)
+  squared_avg = avg ** 2
+  avg_squared = sum([bench**2 for bench in benches])/len(benches)
+  std_dev = (avg_squared - squared_avg) ** 0.5
 
-  return [minimum - diff * RANGE_RATIO_LOWER - avg * ERR_RATIO - ERR_LB,
-          maximum + diff * RANGE_RATIO_UPPER + avg * ERR_RATIO + ERR_UB]
+  # If the results are normally distributed, 2 standard deviations
+  # captures something like ~95% of the possible range of results I think
+  return [avg - 2*std_dev, avg + 2*std_dev]
 
 
 def create_expectations_dict(revision_data_points, builder):