Django Side Effects
Django app for managing external side effects.
This project is now Python3, Django 1.11+ only on master.
The legacy Python2 code is now parked in the python2 branch.
This project was created to try and bring some order to the use of external side-effects within the YunoJuno platform. External side-effects are (as defined by us) those actions that affect external systems, and that are not part of the core application integrity. They fall into two main categories within our application - notifications and updates, and are best illustrated by example:
- HipChat messages
- SMS (via Twilio)
- Push notifications (via Google Cloud Messaging)
- Base CRM (sales)
- Mailchimp CRM (marketing)
- Elasticsearch (full-text index)
There are some shared aspects of all of these side-effects:
- They can all be processed asynchronously (queued)
- They can all be replayed (and are idempotent)
- They can be executed in any order
- They are not time critical
- They do not affect the state of the Django application
As we have continued to build out YunoJuno our use of these side-effects has become ever more complex, and has in some areas left us with functions that are 80% side-effects:
def foo(): # do the thing the function is supposed to do update_object(obj) # spend the rest of the function working out which side-effects to fire if settings.notify_account_handler: send_notification(obj.account_handler) if obj.has_changed_foo(): udpate_crm(obj)
This results in a codebase is:
- Hard to read
- Hard to test
- Hard to document^
^ Barely a week goes by without someone asking "what happens when X does Y - I thought they got email Z?"
This project aims to address all three of the issues above by:
- Removing all side-effects code from core functions
- Simplifying mocking / disabling of side-effects in tests
- Simplifying testing of side-effects only
- Automating documentation of side-effects
It does this with a combination of function decorators that can be used to build up a global registry of side-effects.
The first decorator,
has_side_effects, is used to mark a function as one
that has side effects:
# mark this function as one that has side-effects. The label # can be anything, and is used as a dict key for looking up # associated side-effects functions @side_effects.decorators.has_side_effects('update_profile') def foo(*args, **kwargs): pass
Decorating view functions
By default, the
has_side_effects decorator will run so long as the inner
function does not raise an exception. View functions, however, are a paticular
case where the function may run, and return a perfectly valid
object, but you do not want the side effects to run, as the response object
status_code of 404, 500, etc. In this case, you want to inspect the
inner function return value before deciding whether to fire the side effects
functions. In order to support this, the
has_side_effects decorator has
run_on_exit which takes a function that takes a single parameter,
the return value from the inner function, and must return
which determines whether to run the side effects.
decorators module contains the default argument for this kwarg, a
http_response_check. This will return
False if the
inner function return value is an
HttpResponse object with a status
code in the 4xx-5xx range.
The second decorator,
is_side_effect_of, is used to bind those functions
that implement the side effects to the origin function:
# bind this function to the event 'update_profile' @is_side_effect_of('update_profile') def send_updates(*args, **kwargs): """Update CRM system.""" pass # bind this function also to 'update_profile' @is_side_effect_of('update_profile') def send_notifications(*args, **kwargs): """Notify account managers.""" pass
In the above example, the updates and notifications have been separated out from the origin function, which is now easier to understand as it is only responsible for its own functionality. In this example we have two side-effects bound to the same origin, however this is an implementation detail - you could have a single function implementing all the side-effects, or split them out further into the individual external systems.
Internally, the app maintains a registry of side-effects functions bound to
origin functions using the text labels. The docstrings for all the bound functions can be grouped using these labels, and then be printed out using the
$ ./manage.py display_side_effects This command prints out the first line from the docstrings of all functions registered using the @is_side_effect decorator, grouped by label. update_profile: - Update CRM system. - Notify account managers. close_account: - Send confirmation email to user. - Notify customer service.
If you have a lot of side-effects wired up, you can filter the list by the label:
$ ./manage.py display_side_effects --label update_profile update_profile: - Update CRM system. - Notify account managers.
Or by a partial match on the event label:
$ ./manage.py display_side_effects --label-contains profile update_profile: - Update CRM system. - Notify account managers.
If you want to enforce docstrings on side-effect functions, then you can use the --check-docstrings option, which will exit with a non-zero exit code if any docstrings are missing. This can be used as part of a CI process, failing any build that does not have all its functions documented. (The exit code is the count of functions without docstrings).
$ ./manage.py display_side_effects --check-docstrings update_profile: *** DOCSTRING MISSING: update_crm *** - Notify account managers. ERROR: InvocationError for command '...' (exited with code 1)
Why not use signals?
The above solution probably looks extremely familiar - and it is very closely related to the built-in Django signals implementation. You could easily reproduce the output of this project using signals - this project is really just a formalisation of the way in which a signal-like pattern could be used to make your code clear and easy to document. The key differences are:
- Explicit statement that a function has side-effects
- A simpler binding mechanism (using text labels)
- (TODO) Async processing of receiver functions
It may well be that this project merges back in to the signals pattern in due course - at the moment we still experimenting.
The project is available through PyPI as
$ pip install django-side-effects
And the main package itself is just
>>> from side_effects import decorators
The project has pretty good test coverage (>90%) and the tests themselves run through
$ pip install tox $ tox
If you want to run the tests manually, make sure you install the requirements, and Django.
$ pip install django==2.0 # your version goes here $ tox
If you are hacking on the project, please keep coverage up.
NB If you implement side-effects in your project, you will most likely want to be able to turn off the side-effects when testing your own code (so that you are not actually sending emails, updating systems), but you also probably want to know that the side-effects events that you are expecting are fired.
The following code snippet shows how to use the disable_side_effects context manager, which returns a list of all the side-effects events that are fired. There is a matching function decorator, which will append the events list as an arg to the decorated function, in the same manner that unittest.mock.patch does.
@has_side_effects('do_foo') def foo(): pass def test_foo(): # to disable side-effects temporarily, use decorator with disable_side_effects() as events: foo() assert events = ['do_foo'] foo() assert events = ['do_foo', 'do_foo'] # events list is added to the test function as an arg @disable_side_effects() def test_foo_without_side_effects(events): foo() assert events = ['do_foo']
In addition to these testing tools there is a universal 'kill-switch' which can be set using the env var SIDE_EFFECTS_TEST_MODE=True. This will completely disable all side-effects events. It is a useful tool when you are migrating a project over to the side_effects pattern - as it can highlight where existing tests are relying on side-effects from firing. Use with caution.
Standard GH rules apply: clone the repo to your own account, create a branch, make sure you update the tests, and submit a pull request.
We are using it at YunoJuno, but 'caveat emptor'. It does what we need it to do right now, and we will extend it as we evolve. If you need or want additional features, get involved :-).