Don't fail sum test on near-cancelling inputs

Debian Science Team · rebecca-palmer · commit e7c297e00c63 · 2025-03-29T13:01:52.000Z
Failed in 1.5.3+dfsg-11 after it happened to select an input that
cancelled to ~1e-5 (on i386, so x87 excess precision might be involved)

Author: Rebecca N. Palmer &lt;rebecca_palmer@zoho.com&gt;
Forwarded: no


Gbp-Pq: Name sum_loosen_test_tolerance.patch
diff --git a/pandas/tests/frame/test_reductions.py b/pandas/tests/frame/test_reductions.py
@@ -344,7 +344,7 @@ def sem(x):
             np.sum,
             mixed_float_frame.astype("float32"),
             check_dtype=False,
-            rtol=1e-3,
+            rtol=1e-3, atol=1e-3,
         )
 
         assert_stat_op_calc(

Original file line number	Diff line number	Diff line change
`@@ -344,7 +344,7 @@ def sem(x):`
`344`	`344`	`np.sum,`
`345`	`345`	`mixed_float_frame.astype("float32"),`
`346`	`346`	`check_dtype=False,`
`347`		`- rtol=1e-3,`
	`347`	`+ rtol=1e-3, atol=1e-3,`
`348`	`348`	`)`
`349`	`349`
`350`	`350`	`assert_stat_op_calc(`