How, why and when to use the ".Internal" modules pattern? - haskell

I've seen a couple of package on hackage which contain module names with .Internal as their last name component (e.g. Data.ByteString.Internal)
Those modules are usually not properly browsable (but they may show up nevertheless) in Haddock and should not be used by client code, but contain definitions which are either re-exported from exposed modules or just used internally.
Now my question(s) to this library organization pattern are:
What problem(s) do those .Internal modules solve?
Are there other preferable ways to workaround those problems?
Which definitions should be moved to those .Internal modules?
What's the current recommended practice with respect to organizing libraries with the help of such .Internal modules?

Internal modules are generally modules that expose the internals of a package, that break package encapsulation.
To take ByteString as an example: When you normally use ByteStrings, they are used as opaque data types; a ByteString value is atomic, and its representation is uninteresting. All of the functions in Data.ByteString take values of ByteString, and never raw Ptr CChars or something.
This is a good thing; it means that the ByteString authors managed to make the representation abstract enough that all the details about the ByteString can be hidden completely from the user. Such a design leads to encapsulation of functionality.
The Internal modules are for people that wish to work with the internals of an encapsulated concept, to widen the encapsulation.
For example, you might want to make a new BitString data type, and you want users to be able to convert a ByteString into a BitString without copying any memory. In order to do this, you can't use opaque ByteStrings, because that doesn't give you access to the memory that represents the ByteString. You need access to the raw memory pointer to the byte data. This is what the Internal module for ByteStrings provides.
You should then make your BitString data type encapsulated as well, thus widening the encapsulation without breaking it. You are then free to provide your own BitString.Internal module, exposing the innards of your data type, for users that might want to inspect its representation in turn.
If someone does not provide an Internal module (or similar), you can't gain access to the module's internal representation, and the user writing e.g. BitString is forced to (ab)use things like unsafeCoerce to cast memory pointers, and things get ugly.
The definitions that should be put in an Internal module are the actual data declarations for your data types:
module Bla.Internal where
data Bla = Blu Int | Bli String
-- ...
module Bla (Bla, makeBla) where -- ONLY export the Bla type, not the constructors
import Bla.Internal
makeBla :: String -> Bla -- Some function only dealing with the opaque type
makeBla = undefined

#dflemstr is right, but not explicit about the following point. Some authors put internals of a package in a .Internal module and then don't expose that module via cabal, thereby making it inaccessible to client code. This is a bad thing1.
Exposed .Internal modules help to communicate different levels of abstraction implemented by a module. The alternatives are:
Expose implementation details in the same module as the abstraction.
Hide implementation details by not exposing them in module exports or via cabal.
(1) makes the documentation confusing, and makes it hard for the user to tell the transition between his code respecting a module's abstraction and breaking it. This transition is important: it is analogous to removing a parameter to a function and replacing its occurrences with a constant, a loss of generality.
(2) makes the above transition impossible and hinders the reuse of code. We would like to make our code as abstract as possible, but (cf. Einstein) no more so, and the module author does not have as much information as the module user, so is not in a position to decide what code should be inaccessible. See the link for more on this argument, as it is somewhat peculiar and controversial.
Exposing .Internal modules provides a happy medium which communicates the abstraction barrier without enforcing it, allowing users to easily restrict themselves to abstract code, but allowing them to "beta expand" the module's use if the abstraction breaks down or is incomplete.
1 There are, of course, complications to this puristic judgement. An internal change can now break client code, and authors now have a larger obligation to stabilize their implementation as well as their interface. Even if it is properly disclaimed, users is users and gotsta be supported, so there is some appeal to hiding the internals. It begs for a custom version policy which differentiates between .Internal and interface changes, but fortunately this is consistent with (but not explicit in) the versioning policy. "Real code" is also notoriously lazy, so exposing an .Internal module can provide an easy out when there was an abstract way to define code that was just "harder" (but ultimately supports the community's reuse). It can also discourage reporting an omission in the abstract interface that really should be pushed to the author to fix.

The idea is that you can have the "proper", stabile API which you export from MyModule and this is the preferred and documented way to use the library.
In addition to the public API, your module probably has private data constructors and internal helper functions etc. The MyModule.Internal submodule can be used to export those internal functions instead of keeping them completely locked inside the module.
It lets the users of your libary to access the internals if they have needs that you didn't foresee, but with the understanding that they are accessing an internal API that doesn't have the same implicit guarantees as the public one.
It lets you access the internal functions and constructors for e.g. unit-testing purposes.

One extension (or possibly clarification) to what shang and dflemstr said: if you have internal definitions (data types whose constructors aren't exported, etc.) that you want to access from multiple modules which are exported, then you typically create such an .Internal module which isn't exposed at all (i.e. listed in Other-Modules in the .cabal file).
However, this sometimes does leak out when doing types in ghci (e.g. when using a function but where some of the types it refers to aren't in scope; can't think of an instance where this happens off the top of my head, but it does).

Related

Leave free access to internal intermediary functions in Haskell library?

I'm writing a numerical optimisation library in Haskell, with the aim of making functions like a gradient descent algorithm available for users of the library. In writing these relatively complex functions, I write intermediary functions, such as a function that performs just one step of gradient descent. Some of these intermediary functions perform tasks that no user of the library could ever have need for. Some are even quite cryptic, but make sense when used by a bigger function.
Is it common practice to leave these intermediary functions available to library users? I have considered moving these to an "Internal" library, but moving small functions into a whole different library from the main functions using them seems like a bad idea for code legibility. I'd also quite like to test these smaller functions as well as the main functions for debugging purposes down the line - and ideally would like to test both in the same place, so that complicates things even more.
I'm unsurprisingly using Cabal for the library so answers in that context as well would be helpful if that's easier.
You should definitely not just throw such internal functions in the export of your package's trunk module, together with the high-level ones. It makes the interface/haddocks hard to understand, and also poses problems if users come to depend on low-level details that may easily change in future releases.
So I would keep these functions in an “internal” module, which the “public” module imports but only re-exports those that are indended to be used:
Public
module Numeric.Hegash.Optimization (optimize) where
import Numeric.Hegash.Optimization.Internal
Private
module Numeric.Hegash.Optimization.Internal where
gradientDesc :: ...
gradientDesc = ...
optimize :: ...
optimize = ... gradientDesc ...
A more debatable matter is whether you should still allow users to load the Internal module, i.e. whether you should put it in the exposed-modules or other-modules section of your .cabal file. IMO it's best to err on the “exposed” side, because there could always be valid use cases that you didn't foresee. It also makes testing easier. Just ensure you clearly document that the module is unstable. Only functions that are so deeply in the implementation details that they are basically impossible to use outside of the module should not be exposed at all.
You can selectively export functions from a module by listing them in the header. For example, if you have functions gradient and gradient1 and only want to export the former, you can write:
module Gradient (gradient) where
You can also incorporate the intermediary functions into their parent functions using where to limit the scope to just the parent function. This will also prevent the inner function from being exported:
gradient ... =
...
where
gradient1 ... = ...

Serialization in Haskell

From the bird's view, my question is: Is there a universal mechanism for as-is data serialization in Haskell?
Introduction
The origin of the problem does not root in Haskell indeed. Once, I tried to serialize a python dictionary where a hash function of objects was quite heavy. I found that in python, the default dictionary serialization does not save the internal structure of the dictionary but just dumps a list of key-value pairs. As a result, the de-serialization process is time-consuming, and there is no way to struggle with it. I was certain that there is a way in Haskell because, at my glance, there should be no problem transferring a pure Haskell type to a byte-stream automatically using BFS or DFS. Surprisingly, but it does not. This problem was discussed here (citation below)
Currently, there is no way to make HashMap serializable without modifying the HashMap library itself. It is not possible to make Data.HashMap an instance of Generic (for use with cereal) using stand-alone deriving as described by #mergeconflict's answer, because Data.HashMap does not export all its constructors (this is a requirement for GHC). So, the only solution left to serialize the HashMap seems to be to use the toList/fromList interface.
Current Problem
I have quite the same problem with Data.Trie bytestring-trie package. Building a trie for my data is heavily time-consuming and I need a mechanism to serialize and de-serialize this tire. However, it looks like the previous case, I see no way how to make Data.Trie an instance of Generic (or, am I wrong)?
So the questions are:
Is there some kind of a universal mechanism to project a pure Haskell type to a byte string? If no, is it a fundamental restriction or just a lack of implementations?
If no, what is the most painless way to modify the bytestring-trie package to make it the instance of Generic and serialize with Data.Store
There is a way using compact regions, but there is a big restriction:
Our binary representation contains direct pointers to the info tables of objects in the region. This means that the info tables of the receiving process must be laid out in exactly the same way as from the original process; in practice, this means using static linking, using the exact same binary and turning off ASLR. This API does NOT do any safety checking and will probably segfault if you get it wrong. DO NOT run this on untrusted input.
This also gives insight into universal serialization is not possible currently. Data structures contain very specific pointers which can differ if you're using different binaries. Reading in the raw bytes into another binary will result in invalid pointers.
There is some discussion in this GitHub issue about weakening this requirement.
I think the proper way is to open an issue or pull request upstream to export the data constructors in the internal module. That is what happened with HashMap which is now fully accessible in its internal module.
Update: it seems there is already a similar open issue about this.

What are module signatures in Haskell?

I have recently found Haskell's feature called "module signatures". As I have discovered they are put in .hsig files and begin with signature keyword instead of module.
The example syntax of such a file may look like
signature Str where
data Str
empty :: Str
append :: Str -> Str -> Str
However, I cannot imagine how and why one would use them. Could you explain me which problems do they solve and how to properly make use of them?
They strongly remind me the module system that one can see in OCaml (link), which also has modules signatures and separate implementations, but I can't decide how close are these two concepts. Is it somehow related?
They are related to the OCaml module system, but with some important differences:
Signatures are defined within the language (in .hsig files) but unlike in OCaml they are not instantiated within the language. Instead, the package manager controls instantiation (currently, only Cabal provides that). Modules never know if they are importing an abstract signature or an actual module.
Implementation modules know nothing about signatures and do not not reference them directly. Any existing module can implement a signature if the definitions happen to be compatible.
Instantiation is triggered by a coincidence of module name and signature name in the dependencies of some compilation unit (executable, library, test suite...) When the names coincide, a process called "signature matching" takes place that verifies that the types and definitions are compatible.
The "happy path" is that in your program you depend on some library having a signature "hole", and also on another library that provides an implementation module with the same name. Then signature matching happens automatically. When the names don't match, or we need multiple instantiations of the signature-using library, we have to rename signatures and/or modules in the mixins section of the Cabal file.
As for why module signatures might be useful, consider bytestring, the most popular library by far for handling binary data in Haskell. But there are others, for example stdio with its Bytes type.
Suppose you are writing your own library that uses binary data, and you don't want to force your users into either stdio or bytestring. What are your choices?
One would be to create something like a Bytelike class and parameterize all your functions with it. You would also need to add a type parameter to every data type that contains bytes.
Another would be to create a signature that defines an abstract binary data type and all the operations that are required of it. Your library would make use of the signature, and remain "indefinite" until the user depends both on your library and a suitable implementation when creating his own libraries.
From the perspective of the user, the typeclass solution is unsatisfactory. The user knows that he wants to use either ByteString or Bytes, just one of them. The decision will not depend on some runtime flag and will remain constant across the extent of his program. And yet he has to deal with a more complex API that reminds him of that already decided issue at every turn.
It's better if he makes the decision once, writes it in his .cabal file, and deals with a simpler API from then onwards.
As described here, they're quite closely related to OCaml module signatures. They allow you to create a package missing some modules and say these modules, containing such and such types and values, should be delivered by package's user. I haven't tested it myself, but I imagine that such a package works very much like an OCaml functor.

Puppet Include vs Class and Best Practices

When should I be using an include vs a class declaration? I am exploring creating a profile module right now, but am struggling with methodology and how I should lay things out.
A little background, I'm using the puppet-labs java module which can be found here.
My ./modules/profile/manifests/init.pp looks like this:
class profile {
## Hiera Lookups
$java_version = hiera('profile::jdk::package')
class {'java':
package => $java_version,
}
}
This works fine, but I know that I can also remove the class {'java': block of the code and instead use include java. My question relates around two things. One, if I wanted to use an include statement for whatever reason, how could I still pass the package version from hiera to it? Second, is there a preferred method of doing this? Is the include something I really shouldn't be using, or are there advantages and disadvantages to each method?
My long term goal will be building out profile like modules for my environment. Likely I would have a default profile that applies to all of my servers, and then profiles for different application load outs. I could include the profiles into a role and apply things to my individual nodes at that level. Does this make sense?
Thanks!
When should I be using an include vs a class declaration?
Where a class declares another, internal-only class that belongs to the same module, you can consider using a resource-like class declaration. That leverages your knowledge of the implementation details of the module, as you need to be able to prove that no other declaration of the class in question will be evaluated before the resource-like one. If ever that constraint is violated, catalog building will fail.
Under all other circumstances, you should use include or one of its siblings, require and contain.
One, if I wanted to use an include statement for whatever reason, how
could I still pass the package version from hiera to it?
Exactly the same way you would specify any other class parameter via Hiera. I already answered that for you.
Second, is
there a preferred method of doing this?
Yes, see above.
Is the include something I
really shouldn't be using, or are there advantages and disadvantages
to each method?
The include is what you should be using. This is your default, with require and contain as alternatives for certain situations. The resource-like declaration syntax seemed good to the Puppet team when they first introduced it, in Puppet 2.6, along with parameterized classes themselves. But it turns out that that syntax introduced deep design problems into the language, and it has been a source of numerous bugs and headaches. Automatic data binding was introduced in Puppet 3 in part to address many of those, allowing you to assign values to class parameters without using resource-like declarations.
The resource-like syntax has the single advantage -- if you want to consider it one -- that the parameter values are expressed directly in the manifest. Conventional Puppet wisdom holds that it is better to separate data from code, however, so as to avoid needing to modify manifests as configuration requirements change. Thus, expressing parameter values directly in the manifest is a good idea only if you are confident that they will never change. The most significant category of such cases is when a class has read data from an external source (i.e. looked it up via Hiera), and wants to pass those values on to another class.
The resource-like syntax has the great disadvantage that if a resource-like declaration of a given class is evaluated anywhere during the construction of a catalog for a given target node, then it must be the first declaration of that class that is evaluated. In contrast, any number of include-like declarations of the same class can be evaluated, whether instead of or in addition to a resource-like declaration.
Classes are singletons, so multiple declarations have no more effect on the target node than a single declaration. Allowing them is extremely convenient. Evaluation order of Puppet manifests is notoriously hard to predict, however, so if there is a resource-like declaration of a given class somewhere in the manifest set, it is very difficult in the general case to ensure that it is the first declaration of that class that is evaluated. That difficulty can be managed in the special case I described above. This falls into the more general category of evaluation-order dependencies, and you should take care to ensure that your manifest set is free of those.
There are other issues with the resource-like syntax, but none as significant as the evaluation-order dependency.
Clarification with respect to automated data binding
Automated data binding, mentioned above, associates keys identifying class parameters with corresponding values for those parameters. Compound values are supported if the back end supports them, which the default YAML back end in fact does. Your comments on this answer suggest that you do not yet fully appreciate these details, and in particular that you do not recognize the significance of keys identifying (whole) class parameters.
I take your example of a class that could on one hand be declared via this resource-like declaration:
class { 'elasticsearch':
config => { 'cluster.name' => 'clustername', 'node.name' => 'nodename' }
}
To use an include-like declaration instead, we must provide a value for the class's "config" parameter in the Hiera data. The key for this value will be elasticsearch::config (<fully-qualified classname> :: <parameter name>). The associated value is wanted puppet-side as a hash (a.k.a. "associative array", a.k.a. "map"), so that's how it is specified in the YAML-format Hiera data:
elasticsearch::config:
"cluster.name": "clustername"
"node.name": "nodename"
The hash nature of the value would be clearer if there were more than one entry. If you're unfamiliar with YAML, then it would probably be worth your while to at least skim a primer, such as the one at yaml.org.
With that data in place, we can now declare the class in our Puppet manifests simply via
include 'elasticsearch'

Benefit of importing specific parts of a Haskell module

Except from potential name clashes -- which can be got around by other means -- is there any benefit to importing only the parts from a module that you need:
import SomeModule (x, y, z)
...verses just importing all of it, which is terser and easier to maintain:
import SomeModule
Would it make the binary smaller, for instance?
Name clashes and binary size optimization are just two of the benefits you can get. Indeed, it is a good practice to always identify what you want to get from the outside world of your code. So, whenever people look at your code they will know what exactly your code requesting.
This also gives you a very good chance to creat mocking solutions for test, since you can work through the list of imports and write mockings for them.
Unfortunately, in Haskell the type class instances are not that easy. They are imported implicitly and so can creates conflicts, also they may makes mocking harder, since there is no way to specify specific class instances only. Hopefully this can be fixed in future versions of Haskell.
UPDATE
The benifits I listed above (code maintenance and test mocking) are not limited to Haskell. Actually, it is also common practice in Java, as I know. In Java you can just import a single class, or even a single static variable/method. Unfortunately again, you still cannot selectively import member functions.
No, it's only for the purpose of preventing name clashes. The other mechanism for preventing name clashes - namely import qualified - results in more verbose (less readable) code.
It wouldn't make the binary smaller - consider that functions in a given module all reference each other, usually, so they need to be compiled together.

Resources