Should Domain Model Classes always depend on primitives?
I wrote the book, so I can at least have a go at answering your question.
You can use things other than primitives (str, int, boolean etc) in your domain model. Generally, although we couldn't show it in the book, your model classes will contain whole hierarchies of objects.
What you want to avoid is your technical implementation leaking into your code in a way that makes it hard to express your intent. It would probably be inappropriate to pass instances of Numpy arrays around your codebase, unless your domain is Numpy. We're trying to make code easier to read and test by separating the interesting stuff from the glue.
To that end, it's fine for you to have a DepthMap class that exposes some behaviour, and happens to have a Numpy array as its internal storage. That's not any different to you using any other data structure from a library.
If you've got data as a flat file or something, and there is complex logic involved in creating the Numpy array, then I think a Factory is appropriate. That way you can keep the boring, ugly code for producing a DepthMap at the edge of your system, and out of your model.
If creating a DepthMap from a string is really a one-liner, then a classmethod is probably better because it's easier to find and understand.
I think it's perfectly fine to depend on librairies that are pure language extensions or else you will just end up with having to define tons of "interface contracts" (Python doesn't have interfaces as a language construct -- but those can be conceptual) to abstract away these data structures and in the end those newly introduced contracts will probably be poor abstractions anyway and just result in additional complexity.
That means your domain objects can generally depend on these pure types. On the other hand I also think these types should be considered as language "primitives" (native may be more accurate) just like
datetime and that you'd want to avoid primitive obsession.
In other words,
DepthMap which is a domain concept is allowed to depend on
Numpy for it's construction (no abstraction necessary here), but
Numpy shouldn't necessarily be allowed to flow deep into the domain (unless it's the appropriate abstraction).
Or in pseudo-code, this could be bad:
Where this may be better:
class DepthMap(Numpy: data); someOperation(DepthMap depthMap);
And regarding the second question, from a DDD perspective if the DepthMap class has a Numpy array as it's internal structure but has to be constructed from other sources (string or list for example) would the best approach be a repository pattern? Or is this just for handling databases and a Factory is a better approach?
The Repository pattern is exclusively for storage/retrieval so it wouldn't be appropriate. Now, you may have a factory method directly on
DepthMap that accepts a
Numpy or you may have a dedicated factory. If you want to decouple
Numpy then it could make sense to introduce a dedicated factory, but it seems unnecessary here at first glance.