1. Simplicity

• Write sample client code in the README:
  • E.g. Pendulum's README starts with sample code.
• Reduce boilerplate code: count the number of lines up to first really useful API call:
  • E.g. urllib2 has lots of boilerplate code to do an HTTP request compared to Requests library.
• Document use cases:
  • E.g. python-social-auth document common use-cases.
• Practice Progressive disclosure:
  • Have default arguments and decide their values according to the most common use cases.
  • Have default settings and decide their values according to the most common use cases.
  • Decide argument order by listing the most used ones first and grouping related ones together.
    • E.g. JavaScript history.pushState has bad default argument order: state, title, URL. Most API clients just want to add a URL to the history and they're forced to specify state and title.
  • Don't require copy and pasting of code snippets.
• Avoid cumbersome inputs:
  • Check if parameter name is misleading.
    • E.g. In Scrapy 1.2, the send method accepts a "to" parameter as a list of strings. If the client passes a single string, the method will iterate over the string, trying to send emails to each character of it. Scrapy 1.3 fixed this by accepting both a single string and a list of strings.
  • Check if the user is instantiating something just to call the API. If so, consider accepting the wrapped value.
    • E.g. a function that only accepts file-like objects will force clients to use StringIO if they want to pass strings.
  • Check if it's possible to replace a custom type with a built-in one, or support both.
• Respect the Principle of least astonishment:
  • "If a feature has a high astonishment factor, it may be necessary to redesign it".
  • The default behavior does what clients expect?
  • The default behavior matches client expectations about performance, side-effects, safety, security, restrictions, etc?
• Create abstractions:
  • Draw away from the physical HOW nature of the solution to focus on the WHAT nature.
    • E.g. to make a simple Celery task, devs don't need to worry much about task queues, workers, message brokers, serialization, etc. They just need to use the @app.task decorator. The API is focused on the task definition, not on the tasks inner-workings.
  • Is the API abstracting something it shouldn't be? Be careful with The Law of Leaky Abstractions.
    • E.g. RPC is almost naturally bad because it abstracts remote resources as local ones, but those should really be handled differently from local ones.
• Be Pythonic:
  • Strive for common Python idioms, try to make API calls look like built-in Python API calls.
    • E.g. in Python 2, ConfigParser.get receives a section and option. This isn't natural in Python since get of dict receives key and default value. In Python 3, this was fixed by introducing a dict-like interface to ConfigParser.

2. Consistency

• Naming: is there naming consistency between the API and the Python ecosystem?
  • Follow PEP8, the official style guide for Python code. To guarantee adherence to Naming Conventions and other style constraints, validate the API code with flake8.
• Naming: is there naming consistency inside the API?
  • Order of terms: string_encode vs decode_string.
  • Abbreviations: activate_prev vs fetch_previous, bin2hex vs strtolower.
  • With or without underscore: gettype vs get_class.
  • Plurals: values_list vs value_list.
  • Negatives: button.enabled == True vs button.disabled == True.
• Emptiness: is there consistency in the meaning of the empty value? Is it the best representation?
  • Decide what will be used as empty: None, False, [], '', 0, etc.
  • Be careful with unexpected falsy values: bool(datetime.time(0)) == False in Python < 3.5.
• Parameters: is there consistency in the order of parameters?
  • E.g. datetime.datetime(year, month, day, minute, second, microsecond) vs datetime.timedelta(days, seconds, microseconds, milliseconds, minutes, hours, weeks).
• Behavior: is there consistency in the representation of similar and different behaviors? Asymmetry of behavior should reflect in asymmetry of form:
  • E.g. Good: numbers.sort() in-place vs sorted(numbers) not-in-place.

3. Flexibility:

• Reduce integration discontinuities:
  • Check if classes have a single responsibility. If not, break them into multiple classes:
    • E.g. a class to get/set items from/to a cache should separate the behavior to connect to the cache server into another class.
  • Check if function name has "and" or implies multiple operations. If so, consider breaking into multiple functions. However, keep the combined one if it's used by most clients:
    • E.g. print_formatted function should be broken into two: print and formatted.
  • Check if the API client has to copy-and-paste code to change some behavior. Provide hooks, methods to be overridden, instead.
  • Check if an attribute used internally in class behavior could be a get_something method:
    • E.g. on Django REST Framework, the CursorPagination class supported only a fixed page_size, because it didn't have the get_page_size method. This was changed by introducing the method.
  • Avoid hiding parameters that could be useful:
    • E.g. the API is calling another lower-level one, but it isn't exposing some useful parameters the lower-level API supports.
  • Return everything the API client might need.
  • Pass everything the API client might need when calling it's code (when Inversion of Control is in place).
  • Evaluate every mock.patch in the API tests. This indicates something isn't flexible to be changed at runtime. Consider having parameters to change those things.
    • E.g. Python's built-in sched.scheduler accepts timefunc and delayfunc, so the tests don't need to mock time.monotonic nor time.sleep and the clients can change those behaviors.
• Create abstractions:
  • Wrap resources and functionalities common to multiple low-level implementations.
    • E.g. Beautiful Soup provides the same API for multiple parsers.
  • Provide multiple abstraction levels, from the most simple to the most customizable:
    • E.g. Celery supports both the @app.task decorator and a custom task class that inherits from celery.Task.
  • Provide the ability to get out of the abstraction cage and use the abstracted resources and functionalities directly:
    • E.g. Django querysets support .extra to combine custom SQL with ORM-generated one. It also supports .raw, which allow completely raw SQL queries.
  • Wrap errors that are common to multiple low-level implementations and avoid leaking them directly to the API client:
    • E.g. a database connection error should be the same if the API supports multiple database engines. This helps API clients to change engines without changing much code.
• Be Pythonic:
  • Use properties for simple getters and setters.
  • Use magic methods for operator overloading.
  • Use __repr__ magic method for simpler debug.
  • Use with-statement contexts for open-close/start-end lifecycle.
  • Use decorators to easily compose common behaviors or register something.
  • Use iterators for lazy iterables.
  • Use generators to intertwine logic.
  • Use asyncio for async code.
  • Use built-in collections instead of custom repositories.
  • Accept simple high-order functions for simple Inversion of Control:
    • E.g. list.sort accepts key as the ranker function to calculate the order.
  • For complex flows, consider function/class pipelines, inheritance or generators, in that order:
    • E.g. pipelines: python-social-auth pipelines, inheritance: Django class-based views, generators: Scrapy spiders.
  • Respect duck-typing and "Easier to ask for forgiveness than permission".
• Internationalize (I18N) the strings the end-user will see.

4. Safety

• Check for warning words like "warning", "careful", "remember to", "don't forget" in docs while describing a behavior. If they are there, consider changing behavior to be safer.
• Detect common mistakes and use Python's built-in warnings module to log warnings about them.
• Require unsafe behavior to be explicit:
  • E.g. Don't show all if something isn't set. fields = None shouldn't mean all fields.
• Don't connect code implicitly by name or module. Use a registry or a registry-decorator:
  • E.g. django-admin registry, which supports both register by function or by decorator.
• Avoid depending on order of method calls, try to use with-statement contexts instead.
• Fail-fast: crap in crap out is not a Pythonic idea:
  • Raise ValueError if the library receives an invalid argument, like something that overflows, has a wrong format or is in a wrong state.
  • Raise TypeError if the library receives an argument with incompatible type, like a duck that doesn't quack. But don't do it inside a if isinstance(duck, LibDuck) or if type(duck) == LibDuck)! First try to call quack, then raise TypeError if it fails to give a clearer error.

5. Conclusion

• My API tries to make the simple easy, the complex possible and the wrong impossible.