Get the code running under Valgrind

[tobson]

When running the full test suite, we sometimes encounter the following error.

> assert all(self["y"][:self.np] >= self.manifold.edges[0])
E AssertionError

It's not entirely clear what's causing this error, but it probably has something to do with uninitialized memory. In order to catch errors like this, it would be extremely helpful to get the code running under Valgrind. For a Python code this is non-trivial but possible, see http://scikit-learn.org/stable/developers/debugging.html

[tobson]

I just had this happen again. Both times the failing test was the two=stream instability test.

[tobson]

Currently running the two-stream instability test over and over again, I get this error almost every other time. The good news is that it's evidently fairly easy to reproduce. That should make it a lot easier to track down. Let's try to squash this bug as soon as we can.

[tobson]

I have not seen any assertion errors since this commit: ecca866

[tobson]

I would find it a little bizarre if the change in ecca866 were really the cure, though. Why does the error happen half way through the run as opposed to, say, immediately?

[tobson]

However bizarre I may find this, the change in ecca866 really does seem to be the cure. I ran the the two-stream test 20 times without error. By comparison, repeating the excercise with the previous version 24d0116, the test failed 4 out of 20 times.

Before we stop worrying about this, though, let's make sure we really understand what was going on.

[tobson]

Here's some more info on debugging Python code with Valgrind:

• http://pythonsweetness.tumblr.com/post/151938194807/debugging-c-extension-memory-safety-with-valgrind
• http://stackoverflow.com/a/20126802