# C[omp]ute

Welcome to my blog, which was once a mailing list of the same name and is still generated by mail. Please reply via the "comment" links.

Always interested in offers/projects/new ideas. Eclectic experience in fields like: numerical computing; Python web; Java enterprise; functional languages; GPGPU; SQL databases; etc. Based in Santiago, Chile; telecommute worldwide. CV; email.

© 2006-2013 Andrew Cooke (site) / post authors (content).

## Dynamic Dispatch in Python

From: andrew cooke <andrew@...>

Date: Mon, 9 May 2011 17:51:36 -0400

This is an example of how the pytyp type library can be useful.  It's a piece
of code from another part of the pytyp library that rewrites JSON data as
Python classes.

IMPORTANT: I don't expect anyone to understand the code below.  That's not the

The code below is a complex function.  It contains a lot of logic for how to
rewrite different JSON objects.  Lists, for example, should be kept as lists,
but we need to rewrite the contents of the lists (so this rewriting function
needs to call itself recursively).

Hash maps (dicts), however, may be rewritten as Python classes.  That's the
"useful" part of this particular library - it simplifies Python code that
works with JSON.  So when this function finds a dict it needs to compare it
with the specification that the user has entered and do something useful (in
simple terms you can say to the routine "at this point in the JSON data,
convert the data into a class Foo", for example).

Hang on - we're almost at the clever bit...

So this is a complicated function that has to do different things when working
on different types.  Typically in Python you handle a case like this by having
a set of subclasses that all implement the same interface.  And then the
correct action is taken when you call the method on a certain instance.

That's normal OO dispatch - you're chosing a particular method based on the
type of an object.

What pytyp adds here is the ability to dispatch on other function arguments,
instead of "the object before the dot".  So in the call:

my_function(bar, baz)

you can call different functions depending on the type of baz!  For example:

my_function(bar, 2)  =>  my_int_function(bar, 2)

my_function(bar, 'two')  =>  my_str_function(bar, 'two')

When that's applied to this complicated function used to rewrite JSON it lets
us separate out the "routing", which depends on the type of the data, so that
different types are handled by different methods:

class Transcode:

def __call__(self, value, spec):
if isinstance(value, spec):
return value
else:
type_error(value, spec)

@__call__.intercept
def map_instance(self, value:Mapping, spec:Sub(Ins)):
(varargs, varkw, dict_spec) = class_to_dict_spec(spec)
new_value = transcode(value, dict_spec)
args = new_value.pop(Rec.OptKey(varargs), []) if varargs else []
kargs = new_value.pop(Rec.OptKey(varkw), {}) if varkw else {}
args.extend(new_value.pop(index)
for index in sorted(key
for key in new_value.keys()
if isinstance(Rec.OptKey.unpack(key), int)))
kargs.update((Rec.OptKey.unpack(key), value)
for (key, value) in new_value.items())
return spec._abc_class(*args, **kargs)

@__call__.intercept
def other_instance(self, value, spec:Sub(Ins)):
if isinstance(value, spec):
return value
else:
type_error(value, spec)

@__call__.intercept
def sequence(self, value, spec:Sub(Seq)):
return list(spec._for_each(value, lambda c, vsn: (transcode(v, s) for (v, s, n) in vsn)))

@__call__.intercept
def record(self, value, spec:Sub(Rec)):
if spec._int_keys():
return tuple(spec._for_each(value,
lambda c, vsn: (transcode(v, s)
for (v, s, n) in sorted(vsn,
key=lambda vsn: Rec.OptKey.unpack(vsn[2])))))
else:
return dict(spec._for_each(value,
lambda c, vsn: ((n, transcode(v, s)) for (v, s, n) in vsn)))

@__call__.intercept
def alternative(self, value, spec:Sub(Alt)):
def alt(c, vsn):
error = None
for (v, s, _) in vsn:
try:
return transcode(v, s)
except TypeError as e:
error = e
raise error
print(spec, 'is alt!', value)
return spec._for_each(value, alt)

transcode = Transcode()

As I said, you don't need to understand all that.  All you need to see is that
when you call the transcode() function, what is actually called can be any of
those methods above.  The method Transcode.sequence() handles lists of data,
while the method Transcode.map_instance() is the "core" that rewrites maps as
instance of particular classes.  And the method Transcode.__call__() is the
default method that's called if none of the other, specialised methods match
up with the specifications.

Andrew