Please add tests.

Please add tests.

The tests already exist. They failed sometimes. That's how I found it.

Please add tests.

The tests already exist. They failed sometimes. That's how I found it.

Could you show the failure? Define sometimes? Was that failure non-deterministic? I've never seen it...
This should be tested in

Could you show the failure?

[ RUN      ] StatisticsTest.Median
third_party/benchmark/test/statistics_gtest.cc:17: Failure
Expected equality of these values:
  benchmark::StatisticsMedian({1, 2, 3, 4})
    Which is: 3
  2.5

Define sometimes? Was that failure non-deterministic?

1/2 of the time, non-deterministically.

I've never seen it...

If you use libc++ with appropriate debug flags, you will see it. It will randomize [begin, nth) and (nth, end):

https://github.com/llvm/llvm-project/blob/5ec13535235d07eafd64058551bc495f87c283b1/libcxx/include/__algorithm/nth_element.h#L237

A longer sequence will decrease the probability of success factorially.

This should be tested in

benchmark/test/statistics_gtest.cc

Lines 15 to 19 in dfc8a92

	TEST(StatisticsTest, Median) {
	EXPECT_DOUBLE_EQ(benchmark::StatisticsMedian({42, 42, 42, 42}), 42.0);
	EXPECT_DOUBLE_EQ(benchmark::StatisticsMedian({1, 2, 3, 4}), 2.5);
	EXPECT_DOUBLE_EQ(benchmark::StatisticsMedian({1, 2, 5, 10, 10}), 5.0);
	}

.

It is.

Ok, so the actual bug here is iterator invalidation, effectively.
We should dereference center before calling std::nth_element() again, and then everything would still work.
This was really not at all obvious from the patch description, maybe improve it a bit?

But this is still an improvement, so okay.

Ok, so the actual bug here is iterator invalidation, effectively.

No, the iterator is still valid. The problem is an assumption that nth_element(begin, nth, end) doesn't affect (nth, end).

Ok, so the actual bug here is iterator invalidation, effectively.

No, the iterator is still valid. The problem is an assumption that nth_element(begin, nth, end) doesn't affect (nth, end).

I did say "effectively" - if you consider that the center is an iterator
pointing to the median-ish element of the copy.

thanks!