写出简洁的Python代码: 使用Exceptions（转）

add by zhj: 非常好的文章，异常在Python的核心代码中使用的非常广泛，超出一般人的想象，比如迭代器中，当我们用for遍历一个可迭代对象时，

Python是如何判断遍历结束的呢？是使用的StopIteration异常，这点虽然大部分人知道，但如果是让你设计实现Python，我估计一般人不会这样

做。其实异常在Python中使用非常广泛，完全可以代替错误返回码，并不是说有异常不好，这点跟我之前的想法的确不一样，我以前认为异常只是用

来捕获的，很少去raise异常，但自己主动的去raise+catch异常，会使代码更简洁，而且对性能的影响几乎可以忽略，本文作者也提到了这点。在

A tale of two styles一节中，作者对if+返回码，和异常这两种方式进行了比较，使用if和返回码是在做事之前先检查，而使用异常是直接做事，如果

出错了，再去处理。异常与返回码相比，一个优势是：你可以返回到你想返回的那级函数，在那里捕获异常并处理，而return的话，只能是返回到上

一层函数，在上一层函数中处理。很多时候，当函数出错时，我们想返回到最外层的函数去处理，这时用异常就非常方便。

原文：https://jeffknupp.com/blog/2013/02/06/write-cleaner-python-use-exceptions/

Many programmers have had it drilled into their head that exceptions, in any language, should only be used in truly exceptional cases. They're wrong. The Python community's approach to exceptions leads to cleaner code that's easier to read. And that's without the monstrous hit to performance commonly associated with exceptions in other languages.

EDIT: Updated with more useful exception idioms

Using exceptions to write cleaner code?

When I talk about "using exceptions", I'm specifically not referring to creating some crazy exception hierarchy for your package and raising exceptions at every possible opportunity. That will most certainly lead to unmaintainable and difficult to understand code. This notion has been widely discussed and is well summarized on Joel Spolsky's blog.

Note: Python avoids much of the tension of the "error codes vs exceptions" argument. Between the ability to return multiple values from a function and the ability to return values of different types (e.g. None or something similar in the error case) the argument is moot. But this is besides the point.

The style of exception usage I'm advocating is quite different. In short: take advantage of Python built-ins and standard library modules that already throw exceptions. Exceptions are built in to Python at the lowest levels. In fact, I guarantee your code is already using exceptions, even if not explicitly.

Intermezzo: How the for statement works

Any time you use for to iterate over an iterable (basically, all sequence types and anything that defines __iter__() or __getitem__()), it needs to know when to stop iterating. Take a look at the code below:

How does for know when it's reached the last element in words and should stop trying to get more items? The answer may surprise you: the list raises a StopIteration exception.

In fact, all iterables follow this pattern. When a for statement is first evaluated, it calls iter() on the object being iterated over. This creates an iterator for the object, capable of returning the contents of the object in sequence. For the call to iter() to succeed, the object must either support the iteration protocol (by defining __iter__()) or the sequence protocol (by defining __getitem__()).

As it happens, both the __iter__() and __getitem__() functions are required to raise an exception when the items to iterate over are exhausted. __iter__() raises the StopIteration exception, as discussed earlier, and __getitem__() raises the IndexError exception. This is how for knows when to stop.

In summary: if you use for anywhere in your code, you're using exceptions.

LBYL vs. EAFP

It's all well and good that exceptions are widely used in core Python constructs, but why is a different question. After all, for could certainly have been written to not rely on exceptions to mark the end of a sequence. Indeed, exceptions could have been avoided altogether.

But they exist due to the philosophical approach to error checking adopted in Python. Code that doesn't use exceptions is always checking if it's OK to do something. In practice, it must ask a number of different questions before it is convinced it's OK to do something. If it doesn't ask all of the right questions, bad things happen. Consider the following code:

This trivial function is responsible for calling print() on an object. If it can't be print()-ed, it prints an error message.

Trying to anticipate all error conditions in advance is destined for failure (and is also really ugly). Duck typing is a central idea in Python, but this function will incorrectly print an error for types than can be printed but aren't explicitly checked.

The function can be rewritten like so:

If the object can be coerced to a string, do so and print it. If that attempt raises an exception, print our error string. Same idea, much easier to follow (the lines in the try block could obviously be combined but weren't to make the example more clear). Also, note that we're explicitly checking for TypeError, which is what would be raised if the coercion failed. Never use a "bare" except: clause or you'll end up suppressing real errors you didn't intend to catch.

But wait, there's more!

The function above is admittedly contrived (though certainly based on a common anti-pattern). There are a number of other useful ways to use exceptions. Let's take a look at the use of an else clause when handling exceptions.

In the rewritten version of print_object below, the code in the elseblock is executed only if the code in the try block didn't throw an exception. It's conceptually similar to using else with a for loop (which is itself a useful, if not widely known, idiom). It also fixes a bug in the previous version: we caught a TypeError assuming that only the call to str() would generate it. But what if it was actually (somehow) generated from the call to print() and has nothing to do with our string coercion?

Now, the print() line is only called if no exception was raised. If print() raises an exception, this will bubble up the call stack as normal. The else clause is often overlooked in exception handling but incredibly useful in certain situations. Another use of else is when code in the tryblock requires some cleanup (and doesn't have a usable context manager), as in the below example:

How not to confuse your users

A useful pattern when dealing with exceptions is the bare raise. Normally, raise is paired with an exception to be raised. However, if it's used in exception handling code, raise has a slightly different (but immensely useful) meaning.

Here, we have a member function doing some calculation. We want to keep some statistics on how often the function is misused and throws an exception, but we have no intention of actually handling the exception. Ideally, we want to an exception raised in _do_calculation to be flow back to the user code as normal. If we simply raised a new exception from our except clause, the traceback point to our except clause and mask the real issue (not to mention confusing the user). raise on its own, however, lets the exception propagate normally with its original traceback. In this way, we record the information we want and the user is able to see what actually caused the exception.

A tale of two styles

We've now seen two distinct approaches to error handling (lots of ifstatements vs. catching exceptions). These approaches are respectively known as Look Before You Leap (LBYL) and Easier to Ask for Forgiveness than Permission. In the LBYL camp, you always check to see if something can be done before doing it. In EAFP, you just do the thing. If it turns out that wasn't possible, shrug "my bad", and deal with it.

Idiomatic Python is written in the EAFP style (where reasonable). We can do so because exceptions are cheap in Python.

Slow is relative

The fact that the schism over exception usage exists is understandable. In a number of other languages (especially compiled ones), exceptions are comparatively expensive. In this context, avoiding exceptions in performance sensitive code is reasonable.

But this argument doesn't hold weight for Python. There is someoverhead,  of course, to using exceptions in Python. Comparatively, though, it's negligible in almost all cases. And I'm playing it safe by including "almost" in the previous sentence.

Want proof? Regardless, here's some proof. To get an accurate sense of the overhead of using exceptions, we need to measure two (and a half) things:

1. The overhead of simply adding a try block but never throwing an exception
2. The overhead of using an exception vs. comparable code without exceptions
   1. When the exception case is quite likely
   2. When the exception case is unlikely

The first is easy to measure. We'll time two code blocks using the timeitmodule. The first will simply increment a counter. The second will do the same but wrapped in a try/except block.

Here's the script to calculate the timings:

Note that Timer.repeat(repeat=3, number=1000000) returns the time taken to execute the code block number times, repeated repeat times. The Python documentation suggests that the time should be at least 0.2 to be accurate, hence the change to number.
The code prints the best run of executing each code block (LOOP_IF and LOOP_EXCEPT) 10,000,000 times.

Clearly, all we're measuring here is the setup cost of using an exception. Here are the results:

So the presence of an exception increases run time by .3 seconds divided by 10,000,000. In other words: if using a simple exception drastically impacts your performance, you're doing it wrong...

So an exception that does nothing is cheap. Great. What about one that's actually useful? To test this, we'll load the words file found at/usr/share/dict/words on most flavors of Linux. Then we'll conditionally increment a counter based on the presence of a random word. Here is the new timing script:

The only thing of note is the percentage variable, which essentially dictates how likely our randomly chosen word is to be in my_dict.

So with a 90% chance of an exception being thrown in the code above, here are the numbers:

Wow! 3.2 seconds vs 1.3 seconds! Exceptions are teh sux0rz!

If you run them 1,000,000 times in a tight loop with a 90% chance of throwing an exception, then exceptions are a bit slower, yes. Does any code you've ever written do that? No? Good, let's see a more realistic scenario.

Changing the chance of an exception to 20% gives the following result:

At this point the numbers are close enough to not care. A difference of 0.5 * 10^-6 seconds shouldn't matter to anyone. If it does, I have a spare copy of the K&R C book you can have; go nuts.

What did we learn?

Exceptions in Python are not "slow".

To sum up...

Exceptions are baked-in to Python at the language level, can lead to cleaner code, and impose almost zero performance impact. If you were hesitant about using exceptions in the style described in this post, don't be. If you've avoided exceptions like the plague, it's time to give them another look.