Interfaces and Protocols in Python | Hacker News Reader

Interfaces and Protocols in Python

glyph.twistedmatrix.com

130 pointsBerislavLopac40 comments

I’m not trying to start a flame war, but I’m genuinely curious: if you want and need all this, why not use Java?

Java has all the strong typing and static checking, all you need is to type a lot more words.

The key would be that 'you sometimes want it'.

The stuff that `zope.interface` gets you is honestly on the... heavy side of what I think makes sense for a lot of scenarios (especially in Python). Which would explain it's relatively low uptake despite having been around for like... 15-20 years??

I haven't used `typing.Protocol` yet (I am mostly still working in 3.7), but it seems to roughly capture the level of 'heaviness' that seems workable.

I do get your gist though. Sometimes I wonder what I'm doing type annotating stuff.

> I haven't used `typing.Protocol` yet (I am mostly still working in 3.7)

You can use it if you install the typing-extensions package: https://pypi.org/project/typing-extensions/

> I do get your gist though. Sometimes I wonder what I'm doing type annotating stuff.

I have to say, I despise trying to decipher untyped functions. This arg is called "order" - do you want an instance of Order, an order ID, or a function for sorting?

Fine, I'll read the docstring. Instead of using the IDE.

because most ML libraries are python-first...

as for other stuff (web-servers, what not), I don't think python can beat kotlin / typescript / etc

Scripts where startup time matters.

Situations where you don't want the heavy-weight and complex JVM.

Cases where you need to interact with C/C++ code a lot (ML, etc.).

Companies where python is the norm and introducing another language adds a lot of overhead.

GraalVM (and SubstrateVM) is actually fixing a lot of these; I had a short-lived process that I AOT compiled with SubstrateVM, and I spent a lot of time checking I had done it right and not somehow accidentally compiled in my input data as well or something -- I got a two order of magnitude speedup, with fewer than ten lines of code change.

Replace Java with literally anything else (including whatever Java implementation does not have the issue) and the "startup time" red herring is gone.

Then you get different issues including more obscure and untested systems. I responded to someone asking why not use Java. I answered it. Please don't try to change the question after the fact when you don't like the answer.

I know you're trying to pin me as being rude, but no. Your answer is bad and ad-hoc. Just because a language is not Java does not make it magically untested and obscure. Just use Go tbh.

Edit: Okay I see the original guy stated Java. It is what it is. The real useful answer is to just use any other typed language, and not use Python.

Java's punishing startup times is not a red herring, and I'm dumbfounded someone in a forum like HN makes such a claim. Even in serverless applications like AWS Lambda, python and javascript enjoy execution times that are an order of magnitude faster than Java ones, even with hot starts and stuff initialized a priori.

I was saying that it's a red herring to go "oh people don't used typed languages because Java programs start up slow".

Because you don't always need this. In python, it's optional: you can use it when it matters.

Or you could use Kotlin and keep the strong typing and static checking but do away with the "type a lot more words" part ;)

Usually, because you already have a lot of Python code and it's easier to write more of that, than to try and figure out Python <=> Java interop.

Sure in Java you never have this question of structural vs nominal typing. It's all nominal.

Java doesn't support protocols.

A better question is why not use Kotlin? It has most of the things Python has and much more (no decorators _yet_ though, but they are on the roadmap).

The idea of an interface is fundamental to computing. No matter what you do, at the end of the day data exchanged between two systems has to be structed _some how_. E.g., machine code submitted to the CPU, register configurations, C ABI, Python structures, JSON. You _could_ be hand wavy about it, maybe even use machine learning, but then it will just be ambiguous and lead to vulns. IMO the constant insistence to try and find ways around this is a huge setback. I used Python heavily in 2008 and always was annoyed when trying to figure out the essence of an API (which is what you get when there is no concrete interface). Whenever I read a Python codebase, unsurprisingly, it's full of handwaving and the resultant bugs (some people know what they're doing, but the problem is there more than in a typed language).

It's the first time I hear about Zope, and also zope.interface. How widely is this used?

I personally would be hesitant to adopt something depending on some not-well-known library.

Then on the other side, if sth like interfaces are missing in core Python, and typing.Protocol is not quite the same thing, maybe we should try to extend the core typing for sth like typing.Interface? Or maybe first to typing_extensions.

Or, as I understood it, maybe we just need to fix ABCMeta (for MyPy), and then already have what we want?

I'm also not quite sure I understood: Why do I need this base class zope.interface.Interface at all? Why not just implement IPlugin or whatever like it is mentioned there, but not deriving from Interface, and throw NotImplementedError in methods (tools like PyCharm then recognize this is an abstract method). And all concrete types/classes just explicitly need to derive from this IPlugin class. That's mostly what I anyway do already, and what I also see in other projects code. This approach seems to be fine. What's the problem with it?

Python interfaces are an obscure feature. Zope.interface is the the ancestor of Python interfaces.

Which is to say, not widely.

AFAIK I’ve only seen Zenoss (monitoring software) use Zope widely. About the only place where I saw it.

It used to be very very popular, back around python 2.4 days was its peak i think.

That was around 15 years ago roughly from memory.

I was a bit surprised to see zope.interfaces pop up these days. If you are not using the entire Zope Component Architecture I'm not sure how useful it has been since Python 3.

Zope and things build on top of it (Plone CMS e.g.) was fairly popular in the early-mid 2000s and was in my opinion ahead of it's time in many ways. But also overkill for most projects.

Zope was pretty popular a decade ago, never got into it though.

Anyone curious about the history of Zope should read [1] for how the Python web server landscape looked in 2001.

[1] https://www.ukuug.org/events/linux2001/papers/html/SRichter/...

I discovered Zope in late 2000 while looking for an open source web application platform. Somehow I heard about Zope and thus learnt Python version 1.5.

Zope seemed to be on the cusp of greatness. The Zope conference in Paris in early 2001 was my first tech user group meeting. (Was anyone else reading this there?)

I stopped with Zope in 2003. Then 7 years later got a strange sense of deja vu at a bank I joined, where its trading platform used a Python object database possibly inspired by Zope's ZODB.

IMO, protocols are way better than interfaces. They're essentially the same, except you don't need to inherit from a protocol for your object to be compliant.

If you do this, I urge you to add lots of comments explaining what and why you are doing it. All this is relatively new and will be very confusing to the people reading your code.

Actually, zope.interface has been around for 20 years. It just became almost abandoned like the rest of Zope with the era of Django/Flask (and the zope 3 debacle). Nowadays I think Twisted, Pyramid and Plone are the only "famous" projects left using it

zope.interface was never widely adopted.

That is true, but I think there was a point in time (2000-2001?) where Zope was very much the dominant python web framework, and the Zope foundation employed many python core developers.

It's even mentioned in the history of python https://docs.python.org/3/license.html

Zope was always a decade ahead of its time in the Python space and at least two in the general web framework space.

When I had to explain it to clients I joked it suffered from “Martian Technology Syndrome”.

What was the Zope 3 debacle exactly? I have vague memories of something like that from back in the day, but having trouble finding anything concrete.

I honestly don't really know, and would also be curious to find out what happened exactly that made Zope almost disappear

well, it was more or less the same thing as the python3 debacle. Imagine having a big python2 project and all the cool stuff is happening in python3. You can't upgrade without substantial effort (code & testing). Now replace python2 with zope2 and python3 with zope3. You get the idea.

It's also used by certbot, pyjwt, ckan, proxy.py, scrapy, etc. The list is pretty long but I do agree with you that the majority of packages neither use nor should use it.

https://github.com/zopefoundation/Zope/network/dependents?de...

I’d urge lots of comments if you’re using anything that might lead the reader in the direction of Twisted, which has the most impenetrable documentation of any Python library I’ve ever seen.

The O'Reilly book cover for Twisted is a tangled ball of snakes. This seemed like an accurate choice from its reputation.

OMG as a long-time Pyramid user I never thought I would see an article referring to the Zope Component Architecture on the front page of HN.

Thanks for writing that. I need to use Python fairly often, even though I self identify as a Lisp developer. I found the discussion of checking structure with Protocols very useful. I am going to update my current project.

Reply on news.ycombinator.com