Failures seem real. Not sure I understand why getitem is no good, so am not much use as a reviewer...

I wish runtests.py would catch the warnings that cause travis to fail like that. Hopefully fixed now.

Another way the old behavior was bad: np.array(['abc']).astype(bool) fails with a nonsensical error invalid literal for int() with base 10.

This seems like a fairly clear "good move", and code looks fine to me.

As to whether we want a round of deprecation with FutureWarning: I lean towards yes, since it's not a very critical bug, yet the change is quite user-visible and I can even imagine someone depending on it. So I imagine a FutureWarning for 1.14, then change the behavior in 1.15.

One thing that is arguably worse this way is that this no longer works:

>>> np.array([True, False], dtype=bool).astype(str).astype(bool)

but I'm not sure it should work. After all, bool(str(bool(False))) is True in Python.

Edit: (Wait never mind, this errors out in current master)

I lean towards yes, since it's not a very critical bug, yet the change is quite user-visible and I can even imagine someone depending on it.

A naive implementation of adding the following to numpy\core\src\multiarray\arraytypes.c.src:1577

#if defined(IS_TO_BOOL)
            /* Numpy 1.14, 2017-10-11 */
            if (DEPRECATE_FUTUREWARNING(
                "In future, casting string-like types to bool will not cast "
                "to int first. To keep the current behaviour of having strings "
                "of '0's be considered false, use .astype(int).astype(bool)."
                ) < 0) {
                return;
            }
#endif

\leads to the warning being emitted once for every single element in the array!

>>> a = np.array(list('test'))
>>> a.astype(bool)
FutureWarning: In future, casting string-like types to bool will not cast to int first. To keep the current behaviour of having strings of '0's be considered false, use .astype(int).astype(bool).
FutureWarning: In future, casting string-like types to bool will not cast to int first. To keep the current behaviour of having strings of '0's be considered false, use .astype(int).astype(bool).
FutureWarning: In future, casting string-like types to bool will not cast to int first. To keep the current behaviour of having strings of '0's be considered false, use .astype(int).astype(bool).
FutureWarning: In future, casting string-like types to bool will not cast to int first. To keep the current behaviour of having strings of '0's be considered false, use .astype(int).astype(bool).
array([ True,  True,  True,  True], dtype=bool)

Should I only emit the warning if the behaviour will change? That's still pathological for np.array(['0'] * 10000)

Needs rebase, also to get the merged #9856.

Rebased, but not ready to merge without a decision about if and when to FutureWarning

Closing in favor of having discussion in #9875, since this is a wider issue than anticipated