iommi follows a design philosophy that has evolved over several years. It might feel a bit strange in the beginning but we believe it’s well worth it.
We want it to always be possible to create higher abstractions where you can reuse those abstractions with small tweaks without having to change the abstraction to enable this. If you have code that creates a complex page with tables, forms, and help text fragments in several places, then you should be able to reuse that but with a single line of code make a change to a single small detail of that page, like adding a CSS class somewhere.
In standard APIs you often have to copy paste the entire page and make a small change. This hides the difference between the two pages because you spend 99% of the code to say the same thing. Or alternatively you have to pollute the definition of the first page with some super specific option that makes that code worse. We want to avoid both these scenarios.
In short we want to be able to have code that reads like:
It’s like that one, but different like this.
The philosophy has these main parts:
No silent mistakes#
Systems that fail silently are the worst. Like when you try to render a
value in a django template and you just get silence instead of an error telling
you that you misspelled the variable name. Or when you define a
that never gets called because you misspelled “album” (or worse if you renamed
album to record!). Or silent changes to behavior because you upgraded a library
and your overridden method isn’t called because they renamed the function in
the class you inherited from.
In iommi we have an explicit design goal that all mistakes made by the programmer should (if at all possible) be an error with a very good error message telling them how to fix the problem.
To accomplish this we:
don’t override using standard object oriented overriding, instead preferring refining
show you the valid values when you supply an invalid one, one per line and in alphabetical order
detect common mistakes we’ve made and have helpful error messages telling you what to do or why you can’t do that thing
We love Django, but it does silently fail in many places. In iommi we try our hardest to never let you get stuck with silence as your only company.
Everything has a name#
We like to think of GUIs as a tree of items like tables, buttons, links and pages. We want it to be easy to reference an item in this tree so we can change some property of it, ask it about its configuration, or its state, and more. This is why iommi requires names for everything. This might seem overly verbose in the beginning but this is what enables many of the powerful features of iommi and the robust error handling and error messages.
This philosophy is what enables Single point customization with no boilerplate via Namespace dispatching.
Traversing a namespace is done with
. can’t be used in normal python syntax#
If you have a class
Car that has a member
engine of type
let’s say you want to create a
Car with an electric engine. In standard
Car constructor might take an
engine parameter so you’ll end up
with something like:
car = Car(engine=ElectricEngine())
which is fine if you want to replace the entire engine, but if you just wanted to configure a small thing but keep all the defaults this can become noisy:
car = Car( engine=InternalCombustionEngine( turbo=True, cylinders=6, gearbox=SequentialGearbox( clutch_type='double', ), color='blue', doors=4, make='toyota' # ...and on and on!... ) )
Now it’s impossible to see the intent of the programmer: which of all those
options was the single thing they wanted to change and which are copy paste
of the defaults? Turns out in this case it was just the
would like to write:
car = Car(engine.gearbox.clutch_type='double')
but pythons syntax doesn’t allow this. So instead we use
car = Car(engine__gearbox__clutch_type='double')
this is an elegant solution to this problem, one we’ve stolen from Django’s ORM.
Callables for advanced usage, values for the simple cases#
We want the simple cases to be obvious and simple and the complex cases to be possible. To enable this we aim to make it so that every place you can place a value, you can use a lambda. So for example the simple case could be:
form = Form( auto__model=Musician, fields__instrument__initial='guitar', )
but for the more dynamic case we can write:
form = Form( auto__model=Musician, fields__instrument__initial= lambda request, **_: 'guitar' if request.user.is_staff else 'tambourine', )
In this case you have e.g.
field accessible. If you don’t
know which arguments you can use, you can write whatever and you will get an
error message telling you what arguments are available.
Late binding allows us to sometimes avoid doing work, but more importantly it enables us to build more flexible customizations. A concrete example can be to show a column in a table for only staff users even though the table is defined in the module scope, long before there even is a request object.
Late binding is accomplished by two mechanisms:
not creating object structures until the Bind phase
and Callables for advanced usage, values for the simple cases
Declarative/programmatic hybrid API#
decorators enables us to very easily write an API
that can look both like a normal simple python API:
my_table = Table( columns=dict( foo=Column(), bar=Column(), ), sortable=False)
This code is hopefully pretty self explanatory. But the cool thing is that we can do the exact same thing with a declarative style:
class MyTable(Table): foo = Column() bar = Column() class Meta: sortable = False my_table = MyTable()
This style can be much more readable. There’s a subtle different though between the first and second styles: the second is really a way to declare defaults, not hard coding values. This means we can create instances of the class and set the values in the call to the constructor:
my_table = MyTable( columns__foo__include=False, # <- hides the column foo sortable=True, # <- turns on sorting again )
…without having to create a new class inheriting from
the API keeps all the power of the simple style and also getting the
nice syntax of a declarative API.
Prepackaged commonly used patterns (that can still be customized!)#
A pattern you’ll see often in iommi is that we have class methods instead of classes. We call these “shortcuts”. We don’t need to have classes in order to share functionality and in fact we think this hinders composability and hides lack of customizability.
A shortcut is a bunch of config (and sometimes a tiny bit of code) that also
has a name. We use these instead of writing e.g.
Field subclasses. The names of
these shortcuts are also used by the style system to determine what rules to
An important difference between a traditional class and a shortcut is that the config in a shortcut are defaults, not hard behavior. That means we can start with a shortcut that does mostly what we want and then pass one or more arguments to further refine. Again without writing a class.
Single point customization with no boilerplate#
GUIs consists of layers of abstraction like a form containing fields, fields containing input tags, and a button. But in traditional APIs, to customize the input tag of a form field row you must subclass several classes even for very trivial things. Often trivial things also requires copy pasting a template and making a minor change. This leads to lots of code that basically does nothing and it hides the unique and relevant code in the noise of the other cruft around it that is just copy paste or boilerplate.
In iommi we strive to avoid this by enabling one-off customizations with no boilerplate. To set a CSS style on a specific input field inside a form that was automatically generated we can write:
Form( auto__model=Album, fields__year__input__attrs__style__font='helvetica')
See also Everything has a name
Escape hatches included#
It’s frustrating when a library can’t do what you want. But if the library can’t be extended to do what you want it’s even worse. We aim to include escape hatches for when you reach the limits of iommi. You should be able to add your own logic and data without having to subclass or patch the code.
Read the documentation on extra and extra_evaluated for more information.