I think the real point the author is trying to make is: use modular, layered designs of composable components, instead of monolithic APIs that only have a single entry point. The single-entry-point model imposes costs and decisions onto the application that are hard to work around.
I think this is a good point. I think that it's hard to get from there to "more classes are always better" though. More classes don't always make a design more flexible. You have to consciously design for flexibility and reusability, and even then it takes a lot of deep thought and hard work to get your interfaces right, such that they are truly reusable in practice.
Quoting from the end of the article: "And our solution to monolithic code in Python are classes."
So, I guess the question I ask as someone who doesn't write a ton of python code is: Is that true? And if so, why? Is it not possible to compose a layered API that doesn't rely on inheritance / classes in Python?
Joe Armstrong: "I think the lack of reusability comes in object-oriented languages, not in functional languages. Because the problem with object-oriented languages is they've got all this implicit environment that they carry around with them. You wanted a banana but what you got was a gorilla holding the banana and the entire jungle. If you have referentially transparent code, if you have pure functions-all the data comes in its input arguments and everything goes out and leaves no state behind-it's incredibly reusable. You can just reuse it here, there, and everywhere. When you want to use it in a different project, you just cut and paste this code into your new project. Programmers have been conned into using all these different programming languages and they've been conned into not using easy ways to connect programs together. The Unix pipe mechanism-A pipe B pipe C-is trivially easy to connect things together. Is that how programmers connect things together? No. They use APIs and they link them into the same memory space, which is appallingly difficult and isn't cross-language. If the language is in the same family it's OK-if they're imperative languages, that's fine. But suppose one is Prolog and the other is C. They have a completely different view of the world, how you handle memory. So you can't just link them together like that. You can't reuse things. There must be big commercial interests for whom it is very desirable that stuff won't work together."
- Peter Seibel, Coders at Work: Reflections on the Craft of Programming
Something else I want to mention: what's written above will most likely result in some sort of warmed up discussion in regards to object oriented programming versus something else. Or inheritance versus strategies. Or virtual methods versus method passing. Or whatever else hackernews finds worthy of a discussion this time around.
All of that is entirely irrelevant to the point I'm making which is that monolithic pieces of code are a bad idea. And our solution to monolithic code in Python are classes. If your hammer of choice is Haskell then use whatever the equivalent in Haskell looks like. Just don't force me to fork your library because you decided a layered API is not something you want to expose to your user.
1. When OO was on the brink of becoming mainstream, let's say early 90's, there was an awful lot of the most idiotic rubbish talked about reusability. For example I recall one magazine article that, with a completely straight face, told readers that you would be able to buy an "Aeroplane" class that you could slot into pretty much any application that dealt with aeroplanes. So Joe has a point here.
2. This is not an argument for not using classes within a single codebase. And indeed when I see Erlang functions pushing all the relevant state around as parameters in a big mess of tuples and lists like a crazy baglady with all her possessions in a shopping cart, the benefits of encapsulating it all become clear.
No. Don't write classes. Write useful abstractions. It doesn't matter what the abstraction is (a class, a function, or something else you programming language supports)as long as it fits your current architecture and can be easily modified for possible future extensions.(I will agree with the author that one giant monolithic function is probably a bad abstraction).
First, write nothing. If you can solve your problem without writing code (or even better, by deleting code), that is the best solution.
Next, write code which fits your architecture. Sometimes functional composition is the best system for representing your computational structure. Tree operations, for example, are especially amenable to recursive function-based computation.
Sometimes, when you are writing a state machine, for example, an object is the best possible representation. Objects are entirely focused around hidden-implementation finite state machines, and thus mirror your computation almost exactly.
Funny how most people who have only practical training in a handful of languages and no programming language theory at all tend to advocate for the one style of programming that they know well. When all you have is a hammer...
Note: The small paragraph at the end of this article seems to hedge by agreeing with me and essentially calling the reader to disregard what he previously wrote. If he had followed my step one he could have avoided writing the entire article. Think of the complexity saved!
From what you wrote, I'm not sure you read the article. It's not about objects vs functions; it's about hiding a lot of functionality behind an overly-simple interface, with particular attention paid to an example of a parser that takes a whole string rather than a streaming interface.
What annoyed me most was his constant apparent conflation between bytes and characters. But I didn't really see much evidence of treating all the world like a nail.
As Java was mentioned. The Java API is very nice, but there is a layer missing on top. The Java API was written with a early 90s mindset and got most users after 2000. This goes for most Java APIs. IO, Swing, ... The idea is to have LEGO building blocks (BufferedReader) you can plug together. But you need to plug them together all the time. The missing layer e.g. is IO.readIntoString(file). Apache IOUtils, StringUtils etc. fill in this layer for many Java APIs.
The one API that is not powerful enough is Collection. There you have the top layer without the LEGO building blocks. Compare this to the Scala collection API which has the top layers and the building blocks.
For a good API you need both, building blocks to tailer to your specific need (20% of the time) and an easy top layer (80%) to prevent writing the same stuff all the time.
I really don't like the way this was presented, even if I may agree with the idea underneath. It's not about classes at all. It's about better design, but even the examples are strange.
Just from the JSON example:
- Why do I need a class for streaming JSON - Python's got a perfectly good `yield` for returning tokens in such situations.
- Why would I ever design the JSON library to be extendable at the tokenizer level? If you need serialiser / deserialiser, why not just provide a map of types to callbacks / callback as a parameter? Do you really want to extend JSON format itself?
- The 2GB JSON example is just weird. If you care about such use cases, you a) most likely have a limit on data size at webserver level, b) use proper formats for handling that size of data (I really doubt there's no better data representation once you get to GB sizes).
I see his point of view, but he's arguing for one single "hammer" solution, rather than arguing against the monolithic design. His story seems to present some weird back-story really: "I needed to make my data easier to handle, so I started automatically serialising objects into JSON, then they became huge so I have to start streaming them otherwise just parsing of them takes way too long".
> - Why do I need a class for streaming JSON - Python's got a perfectly good `yield` for returning tokens in such situations.
See the msgpack-cli example at the bottom. Say you have a function that returns a generator for tokens in Python. You would need another function that builds objects out of them. How do you customize how objects are being built? A class makes that simpler because each of the methods are extension points you can override.
But yeah, a token stream would be much appreciated.
> - Why would I ever design the JSON library to be extendable at the tokenizer level?
For instance if you want to skip past objects. That's what we're doing for instance for unwanted input data. That implicitly also prevents hash collision DOS attacks because we never build hash tables for things that we don't want. It also gets rid of the suspension of execution when a garbage collector runs and cleans up a nested structure. I can make you tons of examples where the "build a tree, reject a tree" approach destroys an evented Python server.
> - The 2GB JSON example is just weird. If you care about such use cases, you a) most likely have a limit on data size at webserver level, b) use proper formats for handling that size of data (I really doubt there's no better data representation once you get to GB sizes).
Try using a streaming JSON API like Twitter's firehose. Most people just give up and use XML because there is SAX or if they go with JSON they newline delimit it because most libraries are horrible or useless for parsing streamed JSON.
There's a fairly effective rule of thumb: classes--or more accurately objects--are a way to cleanly encapsulate state. If you rely on a mutating state, you probably want an object/class. If not, a function is often better.
Now you can bundle related functions together in a structure, but this structure is morally a module, not an object, let alone a class. Some languages will force you to encode those modules as classes / prototypes / singletons, but that's just a design pattern to circumvent a limitation of the language.
I had originally used this, but then as my API scope extended to more than just using msgpack, having a common API interface for json (i.e. the .loads() and .dumps() method) was found to be more useful.
And while I agree with most of the article, I don't think writing more classes is a one-size-fits-all solution. Classes IMO, only makes sense from a heavily OOP point of view.
I dont think the OP was pointing to OOP when talking about classes. You can have classes without a meaningful class hierarchy that still conforms to the OP's desired programming style using protocols like the iterator protocol.
A better way to phrase the OP's salient point while sidestepping the polemic of OOP is "Standardize and use protocols".
> You can do the same with python's msgpack as the C# version
You can't. The unpacker in msgpack for Python only reads full objects. The C# version lets me go into and out of streaming at will and even skip past objects without reading them into memory. The only thing the Python version of msgpack can do with the unpacker is buffering up bytes internally until an object is ready. That object however could still expand into a 10GB blob of data which I would then have to postprocess.
if classes didn't have such a terribly verbose syntax in basically every language that has them, i'd be less opposed to using them.
in java, c++, or c#, to add a variable to a class, you have to repeat its name 4 times. once to declare, once in the constructor parameters, and once on each side of the assignment. why am i writing the same thing 4 times for what should be a core part of the language?
in haskell, you write it once (i'm not saying haskell's records are nice, but they got that part right). same with rust.
I get what you're saying, but good OO for me usually suggests nullary constructors, letting a class know when something happens as opposed to setting internal state directly, and the class saying let me do something for you using my state instead of giving you my state directly.
Providing a getter at least is sometimes practically unavoidable, so there's 1 repetition, but with a good IDE that's just a quick key combination over the variable name.
I would argue that this repetition is not essential. It happens when you assign a constructor argument to an attribute directly without any modification. But there are case where it doesn't appear:
- the attribute has a different name
- an attribute is set whose value is a function of more than one argument
- an attribute is set to a constant value independent of constructor arguments
In languages requiring attribute declarations the minimum number of occurrences is 2 (declaration and initialization in a constructor). When declarations aren't required then it's just initialization.
How often such circumstances occur is another issue. The form of initialization that you mentioned is probably the most common so languages can provide shortcuts in this case.
> in java, c++, or c#, to add a variable to a class, you have to repeat its name 4 times. once to declare, once in the constructor parameters, and once on each side of the assignment. why am i writing the same thing 4 times for what should be a core part of the language?
Because Java, C++, and C# suck at this. I'm trying to see the relevance to the OP, which is about Python, which rather decidedly lacks this problem. To add a new instance variable in a class, you need to mention it once, or twice if your using a __slots__-based class for compact memory footprint (because then you have to add it to __slots__ and then actually assign it a value somewhere.)
People don't talk about this anymore for some reason, but I think both Jack and Armin are really just approaching API design in 2 different ways - top-down and bottom-up. The problem is, most people stick with the same approach through out and end up ignoring that programmers are mere mortals too, and they have human needs.
Expanding on Armin's dichotomy, top-down designs like Python's open() or jquery plugins start with giving 70-80% of users APIs that are as simple as possible for their most frequent use cases while shielding them from the sausages underneath.
Bottom-up designs like Java's standard library or POSIX start with LEGO building blocks that solve the most fundamental pieces of largely academic computer science problems and just give them out to their end users and expect them to be able to assemble Tower Defense by solving this LEGO puzzle first.
The problem with sticking entirely to their 2 approaches is that you end up either ignoring power users or making healthy adults with normal IQs feel stupid. There is no reason you can't serve 100% of your user base by incrementally evolving your API approaches and provide 2 sets of APIs within the same library, with the top-down easy one shielding the bottom-up sausage factory that takes care of the meat grinding for you. Most API designers don't realize this and won't ever go there. Extremists and egoists with lots of opinions will spending hundreds of man years to promote their One True Way of doing things. They'll say things like "no leaky abstractions!" or "these enterprise people are just making things too complicated to create jobs!", when the simple truth is probably just that they don't understand how people think.
Make your libraries easy to do things that are easy, but make hard things possible too.
Make your libraries easy to do things that are easy, but make hard things possible too.
IME Ronacher follows this dictum very well. I'd suggest that anyone who wants to see this try his Flask and Werkzeug packages.
As for storing extra state, about which many here have complained, I've found it really helpful that I can set werkzeug.http._accept_re to my own RE when I want to do something weird with media types. That is state that the vast number of users won't need to touch, yet the fact that it exists makes life better for someone who does need it. I'm sure there are numerous other examples I haven't had to bother with yet. Would we really be better off if this RE had to be passed in every time we handled a request? (Although I would understand if you argued this value should be stored in an object not in a module. I haven't needed that yet.)
On the other hand, routing to and handling resources with these packages is typically done with functions and decorators only, although the route decorators are methods of the application object. So Ronacher is not any sort of hardass about classes; he just does what works.
IMO, the Flask example is pushing it - considering how large a job rendering a template is, the number of customization points is probably appropriate, but at some point you're going to get lost in a rabbit hole of wrapper functions that call wrapper functions and end up with three equally appropriate levels you might hook into because the code was written to keep you from having to duplicate a single line of code from the library in your alternative implementation-- never mind how confusing that makes things to the casual debugger (who wants to get to the actual meaty code to see what's wrong with it). Flexibility is useful, but it must justify the loss of simplicity.
But yes, in the JSON cases, having more flexibility than a single 'loads' is clearly justified.
Some note: When it comes to debugging I've rarely seen a better stacktrace output than the one Flask presents you in debug mode (provided by Werkzeug). It actually lets you open the damn sourcefile right there. Right now I'm (mis?)using Flask together with CouchDB to build a db application where you can dynamically define your entities, fields, code hooks and so on at runtime and using Flask for this has been great so far. Ronacher's attitude clearly shows through in how easy I could get Flask to behave in ways it was possibly never meant to (dynamically add view definitions, document mappings and so on).
So Armin, if you read this, thank you, it was a joy spending time with your framework (and source code) over the weekend ;)!
> IMO, the Flask example is pushing it - considering how large a job rendering a template is, the number of customization points is probably appropriate, but at some point you're going to get lost in a rabbit hole of wrapper functions that call wrapper functions
The actual implementations do a bit more. I have pretty much overriden every single part of that at one point, if for no other reason than debugging. Some of those hooks were added later because people requested them.
By default, I go with the inclination to put code in functions when I tackle a project, because most of my projects are just products and applications without a public API and testing or debugging the flow in a functional paradigm is much simpler. I can take a function and plug it in the interactive interpreter (in Python) and run it without needing to instantiate other state that's needed for the test.
However, in certain cases, like where I need to write a public API, I have found that having classes as wrappers to the functionality helps it a bit. So really, the "stop-writing-classes-unless-you- absolutely-need-to" guideline still holds true for me.
The discussion on the thread has rotten to much extent, the article doesnt address different languages or paradigms, he just mentions about classes being better than block codes in python. Something more rudimentary, a valuable piece for quick-fix programmers, for hobby programmers. Classes are one way of abstractions.
So he asks for using abstractions instead on no abstraction (block code) but if you have other ways of abstracting code, please go ahead. THis article is good for people with few / no tools, people who are still learning
Seems like he's asking for modularization so we can make special-case adjustments to libraries without completely monkey patching or rewriting them. Seems like a reasonable ask of a mature library.
On the streaming verus resident working set argument... Most of what most programmers deal with doesn't have to scale to deal with huge streaming datasets, so it doesn't get the attention.
I think there is an aspect to this article that won't be understood unless you're really part of the Python community. For a while now, one of Python's selling points to users from other languages has been "you don't need to write all that code" (most likely directed at Java), more specifically you can simply use a function rather than having to define a class just to define static methods etc. Over time, this has grown into the mantra of "You don't need a class just use a function/module". A lot of people seem to follow this viewpoint blindly, so much so that they consider it to be "Pythonic"; the way you "should" write python.
Could not agree more with the idea that people should write more well designed, OO code. This exact same issue is prevalent in JS land and it leads to people saying JS can't be used for real development, when in fact the problem is that the code they write is a mess and that's what's holding them back.
Is there anything to this debate that doesn't boil down to style? Python has a pretty decent object system. It also has first-class, nested functions with closures. You could write good OO code or good functional code according to your taste, and python libraries don't seem to strongly prefer one style over the other. I'm not terribly interested debating OO vs. functional in the general case, but I am interested in the pros and cons of each style in specific contexts.
For example, Python's lack of variable declarations sometimes leads to bugs involving scoping. (I've run into this a couple of times myself). Does this quirk become more of an issue in heavily functional code? Are there other language quirks that become troublesome in heavily OO code? In what circumstances might one style be preferred over another?
In Python, the distinction is less strong than it is in some other languages. Since a class can implement __call__, allowing instances to be directly called like a function, even if an API specifies a function, you can pass a class instance in with a suitable __call__. So using functions doesn't have to tie to to it being a function forever and ever in the future (or breaking reverse compatibility, the way it does in most other languages. This isn't a perfect answer to the objections, but the objection is at its most weak in Python.
[+] [-] haberman|13 years ago|reply
I think this is a good point. I think that it's hard to get from there to "more classes are always better" though. More classes don't always make a design more flexible. You have to consciously design for flexibility and reusability, and even then it takes a lot of deep thought and hard work to get your interfaces right, such that they are truly reusable in practice.
[+] [-] bhntr3|13 years ago|reply
So, I guess the question I ask as someone who doesn't write a ton of python code is: Is that true? And if so, why? Is it not possible to compose a layered API that doesn't rely on inheritance / classes in Python?
[+] [-] contingencies|13 years ago|reply
Joe Armstrong: "I think the lack of reusability comes in object-oriented languages, not in functional languages. Because the problem with object-oriented languages is they've got all this implicit environment that they carry around with them. You wanted a banana but what you got was a gorilla holding the banana and the entire jungle. If you have referentially transparent code, if you have pure functions-all the data comes in its input arguments and everything goes out and leaves no state behind-it's incredibly reusable. You can just reuse it here, there, and everywhere. When you want to use it in a different project, you just cut and paste this code into your new project. Programmers have been conned into using all these different programming languages and they've been conned into not using easy ways to connect programs together. The Unix pipe mechanism-A pipe B pipe C-is trivially easy to connect things together. Is that how programmers connect things together? No. They use APIs and they link them into the same memory space, which is appallingly difficult and isn't cross-language. If the language is in the same family it's OK-if they're imperative languages, that's fine. But suppose one is Prolog and the other is C. They have a completely different view of the world, how you handle memory. So you can't just link them together like that. You can't reuse things. There must be big commercial interests for whom it is very desirable that stuff won't work together."
- Peter Seibel, Coders at Work: Reflections on the Craft of Programming
[+] [-] skrebbel|13 years ago|reply
Something else I want to mention: what's written above will most likely result in some sort of warmed up discussion in regards to object oriented programming versus something else. Or inheritance versus strategies. Or virtual methods versus method passing. Or whatever else hackernews finds worthy of a discussion this time around.
All of that is entirely irrelevant to the point I'm making which is that monolithic pieces of code are a bad idea. And our solution to monolithic code in Python are classes. If your hammer of choice is Haskell then use whatever the equivalent in Haskell looks like. Just don't force me to fork your library because you decided a layered API is not something you want to expose to your user.
[+] [-] tragomaskhalos|13 years ago|reply
1. When OO was on the brink of becoming mainstream, let's say early 90's, there was an awful lot of the most idiotic rubbish talked about reusability. For example I recall one magazine article that, with a completely straight face, told readers that you would be able to buy an "Aeroplane" class that you could slot into pretty much any application that dealt with aeroplanes. So Joe has a point here.
2. This is not an argument for not using classes within a single codebase. And indeed when I see Erlang functions pushing all the relevant state around as parameters in a big mess of tuples and lists like a crazy baglady with all her possessions in a shopping cart, the benefits of encapsulating it all become clear.
[+] [-] russellallen|13 years ago|reply
[+] [-] Uchikoma|13 years ago|reply
[+] [-] 10098|13 years ago|reply
[+] [-] ankitml|13 years ago|reply
[+] [-] Locke1689|13 years ago|reply
First, write nothing. If you can solve your problem without writing code (or even better, by deleting code), that is the best solution.
Next, write code which fits your architecture. Sometimes functional composition is the best system for representing your computational structure. Tree operations, for example, are especially amenable to recursive function-based computation.
Sometimes, when you are writing a state machine, for example, an object is the best possible representation. Objects are entirely focused around hidden-implementation finite state machines, and thus mirror your computation almost exactly.
Funny how most people who have only practical training in a handful of languages and no programming language theory at all tend to advocate for the one style of programming that they know well. When all you have is a hammer...
Note: The small paragraph at the end of this article seems to hedge by agreeing with me and essentially calling the reader to disregard what he previously wrote. If he had followed my step one he could have avoided writing the entire article. Think of the complexity saved!
[+] [-] barrkel|13 years ago|reply
What annoyed me most was his constant apparent conflation between bytes and characters. But I didn't really see much evidence of treating all the world like a nail.
[+] [-] Uchikoma|13 years ago|reply
The one API that is not powerful enough is Collection. There you have the top layer without the LEGO building blocks. Compare this to the Scala collection API which has the top layers and the building blocks.
For a good API you need both, building blocks to tailer to your specific need (20% of the time) and an easy top layer (80%) to prevent writing the same stuff all the time.
[+] [-] dm3|13 years ago|reply
[+] [-] viraptor|13 years ago|reply
Just from the JSON example:
- Why do I need a class for streaming JSON - Python's got a perfectly good `yield` for returning tokens in such situations.
- Why would I ever design the JSON library to be extendable at the tokenizer level? If you need serialiser / deserialiser, why not just provide a map of types to callbacks / callback as a parameter? Do you really want to extend JSON format itself?
- The 2GB JSON example is just weird. If you care about such use cases, you a) most likely have a limit on data size at webserver level, b) use proper formats for handling that size of data (I really doubt there's no better data representation once you get to GB sizes).
I see his point of view, but he's arguing for one single "hammer" solution, rather than arguing against the monolithic design. His story seems to present some weird back-story really: "I needed to make my data easier to handle, so I started automatically serialising objects into JSON, then they became huge so I have to start streaming them otherwise just parsing of them takes way too long".
[+] [-] the_mitsuhiko|13 years ago|reply
See the msgpack-cli example at the bottom. Say you have a function that returns a generator for tokens in Python. You would need another function that builds objects out of them. How do you customize how objects are being built? A class makes that simpler because each of the methods are extension points you can override.
But yeah, a token stream would be much appreciated.
> - Why would I ever design the JSON library to be extendable at the tokenizer level?
For instance if you want to skip past objects. That's what we're doing for instance for unwanted input data. That implicitly also prevents hash collision DOS attacks because we never build hash tables for things that we don't want. It also gets rid of the suspension of execution when a garbage collector runs and cleans up a nested structure. I can make you tons of examples where the "build a tree, reject a tree" approach destroys an evented Python server.
> - The 2GB JSON example is just weird. If you care about such use cases, you a) most likely have a limit on data size at webserver level, b) use proper formats for handling that size of data (I really doubt there's no better data representation once you get to GB sizes).
Try using a streaming JSON API like Twitter's firehose. Most people just give up and use XML because there is SAX or if they go with JSON they newline delimit it because most libraries are horrible or useless for parsing streamed JSON.
[+] [-] willvarfar|13 years ago|reply
I use msgpack in Python (twisted, tornado) precisely because it can consume byte buffers which are not token-aligned.
[+] [-] fab13n|13 years ago|reply
Now you can bundle related functions together in a structure, but this structure is morally a module, not an object, let alone a class. Some languages will force you to encode those modules as classes / prototypes / singletons, but that's just a design pattern to circumvent a limitation of the language.
[+] [-] chewxy|13 years ago|reply
Try this:
I had originally used this, but then as my API scope extended to more than just using msgpack, having a common API interface for json (i.e. the .loads() and .dumps() method) was found to be more useful.And while I agree with most of the article, I don't think writing more classes is a one-size-fits-all solution. Classes IMO, only makes sense from a heavily OOP point of view.
[+] [-] niggler|13 years ago|reply
A better way to phrase the OP's salient point while sidestepping the polemic of OOP is "Standardize and use protocols".
[+] [-] the_mitsuhiko|13 years ago|reply
You can't. The unpacker in msgpack for Python only reads full objects. The C# version lets me go into and out of streaming at will and even skip past objects without reading them into memory. The only thing the Python version of msgpack can do with the unpacker is buffering up bytes internally until an object is ready. That object however could still expand into a 10GB blob of data which I would then have to postprocess.
[+] [-] notallama|13 years ago|reply
in java, c++, or c#, to add a variable to a class, you have to repeat its name 4 times. once to declare, once in the constructor parameters, and once on each side of the assignment. why am i writing the same thing 4 times for what should be a core part of the language?
in haskell, you write it once (i'm not saying haskell's records are nice, but they got that part right). same with rust.
and with a function, you write it once.
[+] [-] benrhughes|13 years ago|reply
[+] [-] dougk16|13 years ago|reply
Providing a getter at least is sometimes practically unavoidable, so there's 1 repetition, but with a good IDE that's just a quick key combination over the variable name.
[+] [-] grn|13 years ago|reply
How often such circumstances occur is another issue. The form of initialization that you mentioned is probably the most common so languages can provide shortcuts in this case.
[+] [-] dragonwriter|13 years ago|reply
Because Java, C++, and C# suck at this. I'm trying to see the relevance to the OP, which is about Python, which rather decidedly lacks this problem. To add a new instance variable in a class, you need to mention it once, or twice if your using a __slots__-based class for compact memory footprint (because then you have to add it to __slots__ and then actually assign it a value somewhere.)
[+] [-] skrebbel|13 years ago|reply
[+] [-] wyuenho|13 years ago|reply
Expanding on Armin's dichotomy, top-down designs like Python's open() or jquery plugins start with giving 70-80% of users APIs that are as simple as possible for their most frequent use cases while shielding them from the sausages underneath.
Bottom-up designs like Java's standard library or POSIX start with LEGO building blocks that solve the most fundamental pieces of largely academic computer science problems and just give them out to their end users and expect them to be able to assemble Tower Defense by solving this LEGO puzzle first.
The problem with sticking entirely to their 2 approaches is that you end up either ignoring power users or making healthy adults with normal IQs feel stupid. There is no reason you can't serve 100% of your user base by incrementally evolving your API approaches and provide 2 sets of APIs within the same library, with the top-down easy one shielding the bottom-up sausage factory that takes care of the meat grinding for you. Most API designers don't realize this and won't ever go there. Extremists and egoists with lots of opinions will spending hundreds of man years to promote their One True Way of doing things. They'll say things like "no leaky abstractions!" or "these enterprise people are just making things too complicated to create jobs!", when the simple truth is probably just that they don't understand how people think.
Make your libraries easy to do things that are easy, but make hard things possible too.
[+] [-] jessaustin|13 years ago|reply
IME Ronacher follows this dictum very well. I'd suggest that anyone who wants to see this try his Flask and Werkzeug packages.
As for storing extra state, about which many here have complained, I've found it really helpful that I can set werkzeug.http._accept_re to my own RE when I want to do something weird with media types. That is state that the vast number of users won't need to touch, yet the fact that it exists makes life better for someone who does need it. I'm sure there are numerous other examples I haven't had to bother with yet. Would we really be better off if this RE had to be passed in every time we handled a request? (Although I would understand if you argued this value should be stored in an object not in a module. I haven't needed that yet.)
On the other hand, routing to and handling resources with these packages is typically done with functions and decorators only, although the route decorators are methods of the application object. So Ronacher is not any sort of hardass about classes; he just does what works.
[+] [-] splicer|13 years ago|reply
[+] [-] comex|13 years ago|reply
But yes, in the JSON cases, having more flexibility than a single 'loads' is clearly justified.
[+] [-] DeepDuh|13 years ago|reply
So Armin, if you read this, thank you, it was a joy spending time with your framework (and source code) over the weekend ;)!
[+] [-] the_mitsuhiko|13 years ago|reply
The actual implementations do a bit more. I have pretty much overriden every single part of that at one point, if for no other reason than debugging. Some of those hooks were added later because people requested them.
[+] [-] cgopalan|13 years ago|reply
However, in certain cases, like where I need to write a public API, I have found that having classes as wrappers to the functionality helps it a bit. So really, the "stop-writing-classes-unless-you- absolutely-need-to" guideline still holds true for me.
[+] [-] ankitml|13 years ago|reply
So he asks for using abstractions instead on no abstraction (block code) but if you have other ways of abstracting code, please go ahead. THis article is good for people with few / no tools, people who are still learning
[+] [-] redangstrom|13 years ago|reply
On the streaming verus resident working set argument... Most of what most programmers deal with doesn't have to scale to deal with huge streaming datasets, so it doesn't get the attention.
[+] [-] jakejake|13 years ago|reply
[+] [-] xentronium|13 years ago|reply
[+] [-] pyeek|13 years ago|reply
I think Armin's post is somewhat in line with the following post from the google testing blog. http://googletesting.blogspot.com/2008/12/static-methods-are...
[+] [-] Offler|13 years ago|reply
[+] [-] cwp|13 years ago|reply
For example, Python's lack of variable declarations sometimes leads to bugs involving scoping. (I've run into this a couple of times myself). Does this quirk become more of an issue in heavily functional code? Are there other language quirks that become troublesome in heavily OO code? In what circumstances might one style be preferred over another?
[+] [-] jerf|13 years ago|reply