ARROW-765: [Python] Add more natural Exception type hierarchy for thirdparty users

[wesm]

I also took the liberty of changing a number of error types in libarrow_python

[jeffknupp]

Isn't this an antipattern (deriving from a built-in exception or, worse, raising them directly)? If a user combines some Arrow code with their own in a try block, they may inadvertently catch, say, an ArrowInvalid Exception when they explicitly were trying to handle a ValueError for non-arrow code in the try block.

tl;dr if you structure your Exceptions like this, client code has no idea from where the issue originated and has no easy way of disambiguating in the except block (if they're even aware of the Arrow Exception class hierarchy to begin with).

[wesm]

Is it an antipattern? I have never thought there was anything wrong either with

• deriving from built-in exceptions, like ValueError
• raising built-in exceptions directly (pretty much all Python data libraries -- pandas, NumPy, scikit-learn, etc. already do this)

So there's a couple ways to go:

• Don't define user-defined exceptions, raise built-in exceptions only
• Don't derive from a base ArrowException -- which is there to help indicate that the error arose in C++ code
• Derive from base Exception (worst option, IMHO)

adding @jreback for more opinions (I did this at his asking)

[jreback]

my main concern was to NOT have the arrow exception hierarchy just be based on Exception. Being able to catch ValueError, TypeError, KeyError, IndexError (subclass) as errors is very pythonic.

As for the naming having ArrowTypeError(ArrowException, TypeError). I am not how useful that actually is in practice. I would say if you have a need to differentiate between two (or more TypeError s), then having a custom exception is useful (as opposed to having to check the actual error message).

So fine with this PR (though maybe raise the standard python exceptions, unless you have a real need for them).

[jreback]

We do have a couple of these in pandas: https://github.com/pandas-dev/pandas/blob/master/pandas/errors/__init__.py

[wesm]

My preference is to only raise these exceptions for errors that originate from the C++ libraries, otherwise use the Python built-in ones. Sound good?

[jreback]

sgtm.