tryme - Error handling for humans

Introduction

Tryme is a module that makes it easier to treat errors as values. Additionally, tryme helps you handle the absence of a value (None) in a more composable way.

The entire tryme library is contained in a single file tryme.py to make it as easy as possible to drop into an existing project or script.

Why should I care?

Treating errors as values makes it easier to retain information related to errors and lets us defer error handling until the last possible moment. In contrast, exceptions are fundamentally lossy in they by convention contain tracebacks and error messages not values. Further, exceptions abort execution by default and break normal code flow.

Here are a few tasks where we might wish to defer error handling.

  • Processing multiple operations in batch where one or more operations may fail
  • Executing long running operations that require multiple retries

Treating failed operations as values rather than exceptions can simplify conditional logic and make programs more composable.

But c’mon exceptions are pythonic

A common convention for handling errors in Python is to raise exceptions. There is no reason however that this is always the best mechanism to handle errors, especially expected errors. The author of this package feels strongly that exceptions are best used to represent exceptional circumstances such as when undefined behavior is encountered.

Error Handling Strategies

A good error handling strategy retains relevant information and defers side effects related to a failure as long as possible to allow upstream code to decide how to handle the failure.

Here are four error handling strategies that I have encountered:

  1. Return a boolean to indicate success/failure
  2. Return a tuple of (err, value) where err represents a possible error and value is well, a value
  3. Return a value for success and raise an exception in the case of an error
  4. Return a value in case of success and a custom Error class that contains information about the failure but is not necessarily a subclass of Exception
  5. Return an instance of Success or Failure containing a value

Strategy #1 Return a boolean

The simplest possible strategy is to return a boolean to indicate the success or failure. For a simple example, let’s build a trivial program to check if Google is accessible.

Our first iteration will simply return a boolean to indicate if Google is accessible.

import requests


def google_is_accessible():
    response = requests.get('http://google.com')
    return response.ok

is_accessible = google_is_accessible()
if is_accessible:
    print('It works!')
else:
    print('Google is not accessible. No idea why')

Awesome! We now can check if Google is accessible from our remote location. The big drawback is that google_is_accessible doesn’t tell me what went wrong in case of a failure nor does give any information to figure that out for myself.

Strategy #2 Return a tuple of (err, value)

A similar approach is to return a tuple of (err, value) where err indicates if an error occurred. This approach is idiomatic in the Go programming language.

import requests


def google_is_accessible():
    response = requests.get('http://google.com')
    return response.ok, response

is_accessible, response = google_is_accessible()
if is_accessible:
    print('It works!')
else:
    print('Google is not accessible, received http status %s' % response.status_code)

This is a big improvement! We now can determine what went wrong should we care.

The main drawback of returning a tuple to indicate errors is that it makes it harder to compose functions. Let’s extend our simple example to try multiple search engines in case Google is not accessible.:

import requests


def duck_duck_go_is_accessible():
    response = requests.get('http://google.com')
    return response.ok, response


def google_is_accessible():
    response = requests.get('http://google.com')
    return response.ok, response


is_accessible, response = google_is_accessible()
if is_accessible:
    print('It works! Using Google')
else:
    print('Google is not accessible, received http status %s. Trying duckduckgo'
          % response.status_code))
    is_accessible, response = duck_duck_go_is_accessible()
    if is_accessible:
         print('It works! Using DuckDuckGo')
    else:
         print('DuckDuckGo is not accessible, received http status %s. Out of options'
               % response.status_code))

The conditionals in the above example can be reduced but it is apparent that returning a tuple adds more conditional logic to your code.

Strategy #3 Raise an Exception

Here is our example using good old try/except:

import requests


class GoogleNotAccessibleError(Exception):
    pass


def google_is_accessible():
    response = requests.get('http://google.com')
    if response.ok:
        return response.ok
    else:
        return GoogleNotAccessibleError(
              "http request to google.com failed with status code %s" % response.status_code)

 try:
     is_accessible = google_is_accessible()
 except GoogleNotAccessibleError as e:
     is_accessible = False
     print(e.message)

There are pros and cons to the above. We do get back some useful information about the failure. Unfortunately, we do not get back the response object so we cannot dig deeper into the response to determine the cause of the error. To get the HTTP status code we have search the error message.

Another drawback is that the raised exception is a side effect that we have to handle immediately and cannot be deferred until later. Raising an exception also generates something we don’t need, a stacktrace.

One big positive here is that we can subclass exception to indicate the particular problem that occurred.

Strategy #4 Return a custom Error in case of Failure

Instead of raising an Exception, you can simply return an Error in case of failure where Error is an object that is an exception or looks a lot like one.

import requests


class GoogleNotAccessibleError():

    def __init__(self, message, response):
        self.message = message
        self.response = response

def google_is_accessible():
    response = requests.get('http://google.com')
    if response.ok:
        return response.ok
    else:
        return GoogleNotAccessibleError(
              "http request to google.com failed", response)

 result = google_is_accessible()
 if result is True:
     print('It worked!')
 else:
     print(result.message)
     print('HTTP request failed with status code %d' result.value.status_code)

This is a big improvement! We can quickly determine if google is accessible and have access to all the information in the request. The main drawback to returning a custom error is that each implementation is likely custom. The calling code has to know the internals of the returned Error class.

Strategy #5 Return an instance of Success or Failure containing a value

This final strategy refines the custom Error with standard semantics. As it turns out there a standard paradigm in the Either class that is present in Clojure, Scala, and other languages. This package implements the Either class under the name Try as your dear author believes it is a more intuitive name.

The Try class has two subclasses, Success and Failure. Success is used to contain the result of an operation that-you guessed it-succeeded. Likewise, Failure contains the result of an operation that failed.

Here is the same task using the Success and Failure:

import requests
from tryme import Success, Failure


def google_is_accessible():
    response = requests.get('http://google.com')
    if response.ok:
        return Success(response)
    else:
        return Failure(response)

 result = google_is_accessible()
 if result.succeeded():
     print('it worked!')
 else:
     response = result.get_failure()
     print('HTTP request failed with status %d' % response.status_code)

We noted earlier that an advantage of returning exceptions is that we can subclass Exception to more specifically indicate the failure. We can do the same with Success in Failure. One obvious omission from our google_is_accessible is that it does not account for a network failure.:

import requests
from tryme import Success, Failure


class ConnectionFailure(Failure):
    pass


def google_is_accessible():
    try:
        response = requests.get('http://google.com')
    except requests.exceptions.ConnectionError as e:
        return ConnectionFailure(e.message)
    if response.ok:
        return Success(response)
    else:
        return Failure(response)

 result = google_is_accessible()
 if result.succeeded():
     print('it worked!')
 elif isinstance(result, ConnectionFailure):
     print(result.get_message())
 else:
     response = result.get_failure()
     print('HTTP request failed with status %d' % response.status_code)

Note that while we return a custom Failure in this case there are many cases where it is quite reasonable to raise an exception. As mentioned earlier, exceptions work well for unexpected behavior and not expected behavior.

Success and Failure have some useful helpers for reporting to the terminal.

The constructors for both Success and Failure take the optional argument message to capture a message intended for the end user. the to_console method prints the message to the terminal if it is not None otherwise prints a a string representation of the contained value.

  • Success.to_console prints the message if set otherwise prints a string representation of the contained value to stdout
  • Failure.to_console prints the message if set otherwise prints a string representation of the contained value to stderr
  • Try.raise_for_error raise an exception if the instance is a Failure
  • Try.fail_for_error if a Failure, print the message and exit with a non-zero return code

Retrying with Style

Let’s say we want to create a single server using a new Cloud computing provider named HighlyVariable Inc. HighlyVariable can provision our new server in a few seconds, several minutes, or occasionally not at all. Your dear author has used cloud services where the “not at all” is not so uncommon an outcome!

Let’s create a server_ready function that returns a Success when the server is ready, a Failure when the operation times out. A “terminal” state such as “Ready” or “Failed” will terminate retries immediately whereas a Failed will continue execution of the server_ready function until 300 seconds after the function was first called.

If our new server is not ready after 300 seconds, server_ready will return an instance of Failure.

from tryme import retry, Success, Failure


def create_server(name):
    return {'Name': name}

status_iterator = iter(['Preparing', 'Preparing', 'Preparing', 'Ready'])


def check_instance_status(name):
    return next(status_iterator)


@retry
def wait_for_server_ready_or_failed(name):
    status = check_instance_status(name)
    if status in ['Ready', 'Failed']:
        return Success(status)
    else:
        return Failure(status)


def server_ready(name):
    # the decorated function will return two values,
    # the result of wrapped function is updated with start and end times of the
    # retry loop and the total count of attempts
    # note that the wrapped value is not modified
    result = wait_for_server_ready_or_failed(name)

    # a failure here only indicates a timeout
    if result.failed():
        return Failure("Server %s not ready after %d seconds and %d attempts"
                       % (name, result.elapsed, result.count))

    # unwrap the value to see what really happened
    status = result.get()
    if status == 'Ready':
        return Success("server %s is ready after %d seconds and %d attempts!"
                       % (name, result.elapsed, result.count))
    else:
        return Failure("server %s failed after %d seconds!"
                       % (name, result.elapsed))

result = server_ready('jenkins')
assert result.succeeded()
print("Server jenkins is ready after %d seconds and %d attempts!"
      % (result.elapsed, result.count))

There something a little weird about the above example. Why did we return Success when the status was “Failed”? This is because the return value of Failure in the wrapped function is a signal to the @retry decorator to continue retrying until the timeout is reached. As noted earlier, you can subclass Success and Failure with names that make more sense for your context. Tryme in fact comes with two subclasses py:class:Stop and py:class:Again. Let’s refactor the previous example to use them.:

from tryme import retry, Success, Failure, Stop, Again

def create_server(name):
    return {'Name': name}

status_iterator = iter(['Preparing', 'Preparing', 'Preparing', 'Ready'])

def check_instance_status(name):
    return next(status_iterator)

@retry
def wait_for_server_ready_or_failed(name):
    status = check_instance_status(name)
    if status in ['Ready', 'Failed']:
        return Stop(status)
    else:
        return Again(status)

def server_ready(name):
    # the decorated function will return two values,
    # the result of wrapped function is updated with start and end times of the
    # retry loop and the total count of attempts
    # note that the wrapped value is not modified
    result = wait_for_server_ready_or_failed(name)

    # a failure here only indicates a timeout
    if result.failed():
        return Failure("Server %s not ready after %d seconds and %d attempts"
                       % (name, result.elapsed, result.count))

    # unwrap the value to see what really happened
    status = result.get()
    if status == 'Ready':
        return Success("server %s is ready after %d seconds and %d attempts!"
                       % (name, result.elapsed, result.count))
    else:
        return Failure("server %s failed after %d seconds!"
                       % (name, result.elapsed))

result = server_ready('jenkins')
assert result.succeeded()
print("Server jenkins is ready after %d seconds and %d attempts!"
      % (result.elapsed, result.count))

Utility methods

The utility method try_out executes a callable and wraps a raised exception in a Failure class. If an exception was not raised, a Success is returned

>>> from tryme import try_out
>>> result = try_out(lambda: 1 / 0)
>>> print(result)  
Failure(ZeroDivisionError('integer division or modulo by zero',))
>>> exc = result.get_failure()
>>> exc 
ZeroDivisionError('integer division or modulo by zero',)
>>> # the following would fail as it does not catch the correct exception, ZeroDivisionError
>>> # result = try_out(lambda: 1 / 0, exception=ValueError)
>>> result = try_out(lambda: 1 / 1)
>>> print(result) 
Success(1)
>>> result.get() 
1

Requirements

  • CPython >= 2.7

Background

This package is inspired by Philip Xu’s excellent monad package. It also takes some inspiration from the excellent vavr library for java and the Scala language. See this excellent tutorial on the Try utility in Scala.

Pssssh! This package uses gasp monads as the core abstraction. Don’t tell anyone! They will sick the procedural programming police on your dear author. While this package does have a Monad class, it does not provide general purpose implementations of monad, applicative, and functor. Further it does not attempt to overload common Python operators to support function composition. This is because your dear author is of the opinion that Python’s syntax is too limited to support monadic composition.

Installation

Install from PyPI:

pip install tryme

Install from source, download source package, decompress, then cd into source directory, run:

make install

License

BSD New, see LICENSE for details.