On 25 Jul 2016, at 12:10, Daniel Sank wrote:

I just randomly clicked through the docs to this: https://twistedmatrix.com/documents/current/api/twisted._threads.IWorker.htm... https://twistedmatrix.com/documents/current/api/twisted._threads.IWorker.htm...

Note that:

1. It is an interface, and I still don't _really_ understand what that means in Twisted.

2. I have no idea what a "task" is. I realize this is python and yay duck-typing but not specifying the expected behavior of an argument seems like a big omission.

So, while the Twisted docs are great, consider not faulting people for being confused/daunted.

Well, at this point I should argue that _threads, being prefixed by an underscore, is technically a private module to Twisted. That means that, realistically, you shouldn’t really need to consult this *at all*: if anyone outside of Twisted is using IWorker then they’re taking their life into their own hands. That largely argues for part 2, though modern Twisted development practice would still require that we document types and interfaces (see https://twistedmatrix.com/documents/current/api/twisted.web._http2.H2Connect... https://twistedmatrix.com/documents/current/api/twisted.web._http2.H2Connect..., for example). As to interfaces, that’s a separate problem. Interfaces in Twisted are documented. See https://twistedmatrix.com/documents/current/core/howto/components.html https://twistedmatrix.com/documents/current/core/howto/components.html for a very lengthy discussion of interfaces in Twisted. Note that Glyph has also written heavily about interfaces: https://glyph.twistedmatrix.com/2009/02/explaining-why-interfaces-are-great.... https://glyph.twistedmatrix.com/2009/02/explaining-why-interfaces-are-great..... Have those documents not helped, or have you been unable to find them? Cory

On Jul 25, 2016, at 3:10 AM, Daniel Sank wrote:

I realize this is not the main point of this thread, but I'd like to make a comment regarding Twisted being opaque.

...

Twisted is open source, so none of the data types are opaque.

That's a non sequitur. A bunch of open source text in a language you don't understand is opaque, or perhaps better called "obscure". Among other things, Twisted's use of interfaces makes the code very hard to understand.

This is not the first time that someone has made this comment, and I find it very discouraging. An interface is a very simple concept - an abstract description of what an object is expected to provide if you're going to do something useful with it. As a specific programmatic instantiation of this concept, zope.interface provides us with two chief advantages, one purely informational and one useful at run-time. The informational one is that rather than saying "this parameter must provide four methods, the first of which is makeConnection, the second of which is (optionally) connectionMade, the third of which is dataReceived, the fourth of which is connectionLost, which take parameters, respectively,..." we can say "@type: L{IProtocol} provider". The programmatic advantage is that we can ask the question directly; not 'do you have attributes with these names' or 'does this function signature match', but IProtocol.providedBy(something), which asks if 'something' even _intends_ to implement all the relevant functionality described by IProtocol. In my view, interface definition is the primary activity of software development; deciding how the pieces of the system fit together and precisely describing what they do. The fact that so many people seem to find either the basic idea of an abstract type, or the concrete instantiation of that idea in the Zope Interface library, so horribly confusing, makes me despair of ever communicating the actually hard stuff that Twisted gets up to in its internals. The main thing that I have heard in the past about what makes interfaces confusing is that people want to trace through the implementation to see what twisted is 'really doing', and the fact that there is -more than one- implementation of a method like 'listenTCP' is the source of the real confusion, underlying the problem with "interfaces". But, that's the whole point: 'listenTCP' is an abstract concept. The better parts of twisted are _more_ abstract, like 'IStreamServerEndpoint.listen', which has dozens of implementations rather than just 3 or 4, or Deferred.callback, which might do anything at all - if the callback chain did one specific thing there would hardly be any point. So, this is more just an emotional appeal from me, than anything the project needs, but I would very much like to understand *what* is so confusing about "interfaces". Is it, as Cory posited, just that the documentation is not properly linked? Or is it that the average Python developer needs a gentle introduction to the entire idea of abstract rather than concrete types? If they do - is it really Twisted's responsibility to provide it to them? Should Zope Interface just have a snazzier website?

...

Twisted is probably the best open source project I have worked with in terms of having documentation which is generated from the code

I agree that the documentation is generally excellent and that the tests coverage is similarly excellent. However, I still find large fractions of the code very hard to comprehend. A while ago I made a serious effort to understand PB and fix some bugs, but the interface stuff combined with some very odd contortions of python class innards eventually lead me to give up. This is despite the friendly helpful attitude of the main developers both here and in IRC (seriously, thanks everyone for your help back then!).

Given that Twisted is often translating network protocol data into Python method calls, one needs both a working domain knowledge of the protocol involved and a robust understanding of Python metaprogramming constructs. It sounds here like where you fell down was mostly in the "Python metaprogramming" area, where PB is especially intense. Unfortunately, framework code just looks like that (the inner guts of Django are similarly, if not more, obscure, for example); it's not especially specific to Twisted. Don't get me wrong, Python metaprogramming _is_ hard. It's something that I know pretty well, but I can recognize that each additional layer of indirection is additional complexity for someone to learn, and metaprogramming is by definition at least 3 layers indirected from your actual problem. But it's also not really specific to Twisted either. This is another case where I'm not sure what to do except to refer people to the language reference and tell them to work through it slowly.

I just randomly clicked through the docs to this: https://twistedmatrix.com/documents/current/api/twisted._threads.IWorker.htm... https://twistedmatrix.com/documents/current/api/twisted._threads.IWorker.htm...

Note that:

1. It is an interface, and I still don't _really_ understand what that means in Twisted.

The interface stuff in Twisted is a wholly separate library, 'zope.interface'. You can read its documentation - both narrative tutorials and API documentation - here: https://docs.zope.org/zope.interface/README.html This is referenced by the Twisted documentation which Cory referred to earlier, but I think it would be helpful to call that out specifically - you can read about interfaces completely separate from Twisted. They don't mean anything special within Twisted itself (beyond our addition of registerAdapter, which is used less and less often in modern Twisted apps).

2. I have no idea what a "task" is. I realize this is python and yay duck-typing but not specifying the expected behavior of an argument seems like a big omission.

Did you miss the part where it said "type: 0-argument callable" in the documentation? The expected behavior of the argument is that it is a thing that can be called, and it takes 0 arguments. Its return value is unspecified because IWorker providers aren't allowed to use its return value.

So, while the Twisted docs are great, consider not faulting people for being confused/daunted.

I don't want to fault people, and there are many issues we can address within Twisted's documentation. For example, culling things from the API documentation and presenting it as it is really intended to be consumed, eliding private methods that have underscores or live in test_* packages to present a smaller surface, improving tutorial documentation, and so on. It's not entirely the user's fault. But certain things - the fact that we might have an abstract interface with multiple concrete implementations, for example - are inherent parts of the problem domain that Twisted is trying to address, and this is what Interfaces let us express. When people have said things like "Twisted's use of interfaces makes -glyph

On Jul 26, 2016, at 12:26 AM, Daniel Sank wrote:

Glyph,

...

...

2. I have no idea what a "task" is. I realize this is python and yay duck-typing but not specifying the expected behavior of an argument seems like a big omission.

...

Did you miss the part where it said "type: 0-argument callable" in the documentation?

Yes :(

I can only guess that I missed it because the type is defined near the end of the description line, whereas I'm used to seeing

def foo(x, y): """Do something.

Args: x (int): blah blah y (banana): yadda yadda """

FWIW, now that I look at the code, the type specification is way more visually apparent there than it is in the generated HTML.

tl,dr: I take it all back and thanks for pointing out the obvious.

This does at least point to a real problem with pydoctor in the way it presents types. It should probably put them in their own colored box, not use the string 'type' or parentheses to offset them, and put the type closer to (rather than farther from) the parameter name. Would you mind filing a bug on pydoctor? Or commenting on one if it already exists? :)

...

An interface is a very simple concept - an abstract description of what an object is expected to provide if you're going to do something useful with it.

Indeed, a general understanding of interfaces is not the problem.

...

In my view, interface definition is the primary activity of software development

Agreed 100%.

OK. Glad to hear it.

...

The fact that so many people seem to find either the basic idea of an abstract type, or the concrete instantiation of that idea in the Zope Interface library, so horribly confusing, makes me despair of ever communicating the actually hard stuff that Twisted gets up to in its internals.

...

I would very much like to understand *what* is so confusing about "interfaces". Is it, as Cory posited, just that the documentation is not properly linked? Or is it that the average Python developer needs a gentle introduction to the entire idea of abstract rather than concrete types? If they do - is it really Twisted's responsibility to provide it to them? Should Zope Interface just have a snazzier website?

Some years ago when I tried to understand Twisted's use of interfaces via Twisted's own documentation (which included something about hair dryers and voltage standards) I was puzzled by the fact that the examples didn't really show me how to solve a useful problem (or I was too stupid to understand that the examples did in fact do that) despite the fact that I knew what an interface was in general terms. It was a case of understanding the intent but none of the examples.

OK... it's a fair cop. That documentation is not the best. Among other things, it's mainly trying to explain adaptation, which sort of puts the cart before the horse, and automatic adaptation is increasingly considered spooky action-at-a-distance within Twisted code. You can see it here: <http://twisted.readthedocs.io/en/latest/core/howto/components.html http://twisted.readthedocs.io/en/latest/core/howto/components.html>. You're the perfect person to submit patches against this doc, by the way, since you have a firm grasp of the whole "abstract interface" thing but also found it confusing. Personally, I find the examples very clear - I say the documentation is "not the best" because I could see how it could confuse somebody _else_, but it doesn't confuse _me_ at all, so it's a bit hard for me to improve it (especially incrementally).

A brief look at the zope documentation just now makes me think the situation has improved.

Well that's good, at least. perhaps we should link to it more prominently.

The other problem was that interfaces were sprinkled somewhat haphazardly around the code I was trying to understand (perspective broker) and it was just plain hard to keep navigating around the code files to understand who was implementing what interfaces. This could have been my own fault for not having a editor set up. I don't know.

Setting up your editor to have a 'jump to definition' key definitely helps; but then, it generally helps with any large codebase. So, hard to say.

...

Should Zope Interface just have a snazzier website?

I think the real issue is the need for compelling and simple examples.

Do you think it would be better to put things in terms of a concrete Twisted interface, like "IProtocol"? I am pretty sure these docs were trying to stay away from anything "real" because this is a highly abstract concept that could apply to anything, and when we drag a concrete example in

P.S. Everything below here is completely off topic of this thread and I probably shouldn't have written it.

In for a penny... (We may want to spin out into a different thread for talking about the PB issue, but it looks like this was unresolved for you, and I can definitely shed some more light.)

...

Given that Twisted is often translating network protocol data into Python method calls, one needs both a working domain knowledge of the protocol involved and a robust understanding of Python metaprogramming constructs. It sounds here like where you fell down was mostly in the "Python metaprogramming" area, where PB is especially intense.

...

But it's also not really specific to Twisted either. This is another case where I'm not sure what to do except to refer people to the language reference and tell them to work through it slowly.

...

Unfortunately, framework code just looks like that

I spent a considerable amount of time reading the PB code, reproducing parts of it myself, and talking to people on IRC and the mailing list to understand a particularly weird issue in PB. See here for the bug I was trying to fix (note in particular my first comment to the one existing answer):

http://stackoverflow.com/questions/23421423/why-are-dummy-objects-created-in... http://stackoverflow.com/questions/23421423/why-are-dummy-objects-created-in...

I distinctly recall that near the end of my efforts you (Glyph) or someone else more or less told me that the PB code was old, horrible, and that the issues I was trying to understand were probably incidental complexity due to poor design etc. You guys were joking around on IRC about how ridiculous all the dummy object construction is. So, I think this particular incident was less due a lack of understanding of python metaprogramming and more due to PB having some bizarre warts.

OK. Maybe I made this sound a bit too simple, but it's still not really Twisted's fault. The bizarre warts here - and they are definitely here - are mostly an outgrowth of the bizarre mismatch between old-style and new-style classes, and the mad shuffle of random API deprecations, often without suitable replacements, or without suitable portable replacements, within the standard library. In the old-style world, you had Class objects. Class objects could be created in a variety of ways, but the Right™ way to make a new, empty class that hadn't had its initializer run was 'new.instance'. Of course, new.instance doesn't work with new-style classes, because now the Right™ way to make a new, empty class that hadn't had its initializer run was yourclass.__new__(). Unless of course you overrode __new__, which is totally allowed, so the caller can't know what that signature is supposed to be. So then you use `object.__new__(yourclass)´ in order to get a known signature - but of course that won't work at all with old-style classes. PB is bridging the gap between these two worlds, and it has to find hacks which work in both and don't draw in any deprecated APIs that are gone in python 3 in order to do it. This means you can't use the Right™ way at all, and instead must resort to contortions which depend on implementation details that happen to be held in common between both new-style and old-style objects. -glyph

On 2016-07-25, 09:55 GMT, Craig Rodrigues wrote: First of all, thank you very much for all help you gave to M2Crypto. However, ...

I call shenanigans on you.

Nothing of which I have been accused of has anything to do with what I meant. When I said „opaque“, I didn’t mean to offend Twisted. Just to say, that while I was marking whole library with PEP-484 type hints, I usually dealt with bytes, str, ints, and very few rather simple objects. With Twisted I get object @implementer(ITLSTransport) class TLSProtocolWrapper(ProtocolWrapper): def __init__(self, factory, wrappedProtocol, startPassThrough, client, contextFactory, postConnectionCheck): # type: policies.WrappingFactory, object, int, int, object, Checker (and with those two objects, I don't even dare to guess what types these are, and all that covered in some weird decorator from Zope (?)). That's nothing wrong with Twisted, just that it is really difficult for idiot like me to understand what's going on.

If you are unfamiliar with Twisted's code and data types, and don't have the energy to dig in, then be honest about that, but don't accuse Twisted of being "opaque", because it isn't.

I don't think what's dishonest on saying that Twisted API is quite complicated and I am stupid enough not understanding what's going on.

Regarding your code example which is failing, your code is failing because you are intermixing bytes and strings which is a big no-no for Python 3.

Of course I know that (it is not the first place where I have to deal with bytes × str dichotomy in py3k), but in order to understand what's going, I have to first understand where these values come from and where they run to, i.e., to decipher Twisted. Thus I was asking for help.

If I look at this line for example: https://gitlab.com/m2crypto/m2crypto/blob/master/M2Crypto/SSL/TwistedProtoco...

I see the code is doing stuff like: data = '' encryptedData = ''

Those are of type str, and need to be of type bytes:

data = b'' encryptedData = b''

You need to clean stuff like that up in your code so that you are only using bytes.

I believe I have fixed all I can do without actually understanding Twisted in https://gitlab.com/mcepl/m2crypto/commit/6cd5f87b31e50016ebb7e44f3f2ae46610b.... So now, if Twisted is so transparent and perfectly understandable, could you please suggest, what I do wrong, that the test ends in the endless loop (https://travis-ci.org/mcepl/M2Crypto/builds/147175901)? Thank you, Matěj Cepl -- https://matej.ceplovi.cz/blog/, Jabber: mcepl@ceplovi.cz GPG Finger: 3C76 A027 CA45 AD70 98B5 BC1D 7920 5802 880B C9D8 Never ascribe to malice that which is adequately explained by stupidity. -- Napoleon Bonaparte (or many other people to whom this quote is ascribed)

On Jul 25, 2016, at 16:02, Matěj Cepl wrote:

On 2016-07-25, 09:33 GMT, Craig Rodrigues wrote:

...

Earlier this year, I contributed lots of patches to you in M2Crypto to port it to py3k. Now I have shifted efforts to Twisted, where in the past month I have contributed hundreds of patches to help improve py3k support in Twisted.

Hi,

can I for a piece of advice on the documentation. M2Crypto.SSL.TwistedProtocolWrapper.TLSProtocolWrapper.startTLS implements ITLSTransport.startTLS interace method, which first parameter is called ``ctx`` and in `the current implementation in M2Crypto`_ it is of type control of SSL Context (that's M2Crypto.SSL.Context.Context) or whether it is factory generating such Contexts (which is what https://twistedmatrix.com/documents/current/api/twisted.internet._newtls.Con... says, at least I understand it in this way). Is my current implementation wrong?

Technically speaking, your implementation is wrong because it claims to implement https://twistedmatrix.com/documents/16.3.0/api/twisted.internet.interfaces.I..., which documents the method startTLS https://twistedmatrix.com/documents/16.3.0/api/twisted.internet.interfaces.I... to accept a `contextFactory´ which is a provider of either https://twistedmatrix.com/documents/16.3.0/api/twisted.internet.interfaces.I... or https://twistedmatrix.com/documents/16.3.0/api/twisted.internet.interfaces.I.... Both of these interfaces return pyOpenSSL-specific objects. If you want to do TLS with M2Crypto, you must therefore give up on supporting any of Twisted's interfaces directly, because (for example) optionsForClientTLS https://twistedmatrix.com/documents/16.3.0/api/twisted.internet.ssl.html#opt... is not going to work with your implementation, since you don't call clientConnectionForTLS on it. If you want to provide TLS purely with M2Crypto, then you should have interfaces which describe exactly how it should work with M2Crypto. You can make it take a concrete context if you want, or a factory, whichever makes sense for how you're going to set it up. Personally my recommendation would be to go in the direction that Twisted itself has been moving and pass a thing that can create Connection objects (or, in OpenSSL-speak, an "SSL*", or in M2Crypto, an _SSLProxy(ssl_new())). No existing Twisted code which is going to call startTLS() can be made to work with these interfaces without extensive monkey-patching, and even then, anything which expects to be able to cut in at the OpenSSL layer will break. Basically, Twisted doesn't have a mechanism for abstracting away the TLS backend yet. I'd really like it if it did! If you want M2Crypto to be able to do what it's currently trying to do, you could contribute code to Twisted to make things like optionsForClientTLS more abstract, and to isolate the TLS implementation more closely to the TLS wrapper factory. This would make it easier to adopt Cryptography's TLS API eventually, which is what we'll need to do as pyOpenSSL eventually becomes less relevant. This is still several years away, of course. But it would be nice to have some help getting there in advance. -glyph