Browse Source

prob-distr: Decrease false positive rate of stochastic tests.

George Kadianakis 5 years ago
parent
commit
a5df9402b6
2 changed files with 12 additions and 6 deletions
  1. 3 0
      changes/bug29693
  2. 9 6
      src/test/test_prob_distr.c

+ 3 - 0
changes/bug29693

@@ -0,0 +1,3 @@
+  o Minor bugfixes (unit tests):
+    - Decrease the false positive rate of stochastic probability distribution
+      tests. Fixes bug 29693; bugfix on 0.4.0.1-alpha.

+ 9 - 6
src/test/test_prob_distr.c

@@ -878,11 +878,14 @@ test_uniform_interval(void *arg)
  * is higher: 1 - Binom(0; n, alpha) = 1 - (1 - alpha)^n.  For n = 10,
  * this is about 10%, and for n = 100 it is well over 50%.
  *
- * We can drive it down by running each test twice, and accepting it if
- * it passes at least once; in that case, it is as if we used Binom(2;
- * 2, alpha) = alpha^2 as the false positive rate for each test, and
- * for n = 10 tests, it would be 0.1%, and for n = 100 tests, still
- * only 1%.
+ * Given that these tests will run with every CI job, we want to drive down the
+ * false positive rate. We can drive it down by running each test four times,
+ * and accepting it if it passes at least once; in that case, it is as if we
+ * used Binom(4; 2, alpha) = alpha^4 as the false positive rate for each test,
+ * and for n = 10 tests, it would be 9.99999959506e-08. If each CI build has 14
+ * jobs, then the chance of a CI build failing is 1.39999903326e-06, which
+ * means that a CI build will break with probability 50% after about 495106
+ * builds.
  *
  * The critical value for a chi^2 distribution with 100 degrees of
  * freedom and false positive rate alpha = 1% was taken from:
@@ -895,7 +898,7 @@ test_uniform_interval(void *arg)
 
 static const size_t NSAMPLES = 100000;
 /* Number of chances we give to the test to succeed. */
-static const unsigned NTRIALS = 2;
+static const unsigned NTRIALS = 4;
 /* Number of times we want the test to pass per NTRIALS. */
 static const unsigned NPASSES_MIN = 1;