Support for well-typed paths.

Path

Support for well-typed paths in Haskell.

• Motivation
• Approach
• Solution
• Implementation
  • The data types
  • Parsers
  • Smart constructors
  • Overloaded stings
  • Operations
• Review
  • Relative vs absolute confusion
  • The equality problem
  • Unpredictable concatenation issues
  • Confusing files and directories
  • Self-documentation
• In practice
• Doing I/O
• Doing textual manipulations
• Accepting user input
• Comparing with existing path libraries
  • filepath and system-filepath
  • system-canonicalpath, canonical-filepath, directory-tree
  • pathtype
  • data-filepath
• Summary

Motivation

It was after working on a number of projects at FP Complete that use file paths in various ways. We used the system-filepath package, which was supposed to solve many path problems by being an opaque path type. It occurred to me that the same kind of bugs kept cropping up:

• Expected a path to be absolute but it was relative, or vice-versa.

• Expected two equivalent paths to be equal or order the same, but they did not (/home//foo vs /home/foo/ vs /home/bar/../foo, etc.).

• Unpredictable behaviour with regards to concatenating paths.

• Confusing files and directories.

• Not knowing whether a path was a file or directory or relative or absolute based on the type alone was a drag.

All of these bugs are preventable.

Approach

My approach to problems like this is to make a type that encodes the properties I want and then make it impossible to let those invariants be broken, without compromise or backdoors to let the wrong value “slip in”. Once I have a path, I want to be able to trust it fully. This theme will be seen throughout the things I lay out below.

Solution

After having to fix bugs due to these in our software, I put my foot down and made:

• An opaque Path type (a newtype wrapper around String).

• Smart constructors which are very stringent in the parsing.

• Make the parsers highly normalizing.

• Leave equality and concatenation to basic string equality and concatenation.

• Include relativity (absolute/relative) and type (directory/file) in the type itself.

• Use the already cross-platform filepath package for implementation details.

Implementation

The data types

Here is the type:

newtype Path b t = Path FilePath
  deriving (Data, Typeable, Generic)

The type variables are:

• b — base, the base location of the path; absolute or relative.
• t — type, whether file or directory.

The base types can be filled with these:

data Abs deriving (Typeable)
data Rel deriving (Typeable)

And the type can be filled with these:

data File deriving (Typeable)
data Dir deriving (Typeable)

(Why not use data kinds like data Type = File | Dir? Because that imposes an extension overhead of adding {-# LANGUAGE DataKinds #-} to every module you might want to write out a path type in. Given that one cannot construct paths of types other than these, via the operations in the module, it’s not a concern for me.)

There is a conversion function to give you back the filepath:

toFilePath :: Path b t -> FilePath
toFilePath (Path l) = l

Beginning from version 0.5.3, there are type-constrained versions of toFilePath with the following signatures:

fromAbsDir  :: Path Abs Dir -> FilePath
fromRelDir  :: Path Rel Dir -> FilePath
fromAbsFile :: Path Abs File -> FilePath
fromRelFile :: Path Rel File -> FilePath

Parsers

To get a Path value, you need to use one of the four parsers:

parseAbsDir  :: MonadThrow m => FilePath -> m (Path Abs Dir)
parseRelDir  :: MonadThrow m => FilePath -> m (Path Rel Dir)
parseAbsFile :: MonadThrow m => FilePath -> m (Path Abs File)
parseRelFile :: MonadThrow m => FilePath -> m (Path Rel File)

The following properties apply:

• Absolute parsers will reject non-absolute paths.

• The only delimiter syntax accepted is the path separator; / on POSIX and \ on Windows.

• Any other delimiter is rejected; .., ~/, /./, etc.

• All parsers normalize into single separators: /home//foo → /home/foo.

• Directory parsers always normalize with a final trailing /. So /home/foo parses into the string /home/foo/.

It was discussed briefly whether we should just have a class for parsing rather than four separate parsing functions. In my experience so far, I have had type errors where I wrote something like x <- parseAbsDir someAbsDirString because x was then passed to a place that expected a relative directory. In this way, overloading the return value would’ve just been accepted. So I don’t think having a class is a good idea. Being explicit here doesn’t exactly waste our time, either.

Why are these functions in MonadThrow? Because it means I can have it return an Either, or a Maybe, if I’m in pure code, and if I’m in IO, and I don’t expect parsing to ever fail, I can use it in IO like this:

do x <- parseRelFile (fromCabalFileName x)
   foo x
   …

That’s really convenient and we take advantage of this at FP Complete a lot. The instances

Equality, ordering and printing are simply re-using the String instances:

instance Eq (Path b t) where
  (==) (Path x) (Path y) = x == y

instance Ord (Path b t) where
  compare (Path x) (Path y) = compare x y

instance Show (Path b t) where
  show (Path x) = show x

Which gives us for free the following equational properties:

toFilePath x == toFilePath y        ≡ x == y           -- Eq instance
toFilePath x `compare` toFilePath y ≡ x `compare` y    -- Ord instance
toFilePath x == toFilePath y        ≡ show x == show y -- Show instance

In other words, the representation and the path you get out at the end are the same. Two paths that are equal will always give you back the same thing.

Smart constructors

For when you know what a path will be at compile-time, there are constructors for that:

$(mkAbsDir "/home/chris")
$(mkRelDir "chris")
$(mkAbsFile "/home/chris/x.txt")
$(mkRelFile "chris/x.txt")

With the QuasiQuotes language extension, paths can be written as follows:

[absdir|/home/chris|]
[reldir|chris|]
[absfile|/home/chris/x.txt|]
[relfile|chris/x.txt|]

These will run at compile-time and underneath use the appropriate parser.

Overloaded strings

No IsString instance is provided, because that has no way to statically determine whether the path is correct, and would otherwise have to be a partial function.

In practice I have written the wrong path format in a $(mk… "") and been thankful it was caught early.

Operations

There is path concatenation:

(</>) :: Path b Dir -> Path Rel t -> Path b t

Get the parent directory of a path:

parent :: Path Abs t -> Path Abs Dir

Get the filename of a file path:

filename :: Path b File -> Path Rel File

Get the directory name of a directory path:

dirname :: Path b Dir -> Path Rel Dir

Stripping the parent directory from a path:

stripProperPrefix :: MonadThrow m => Path b Dir -> Path b t -> m (Path Rel t)

Review

Let’s review my initial list of complaints and see if they’ve been satisfied.

Relative vs absolute confusion

Paths now distinguish in the type system whether they are relative or absolute. You can’t append two absolute paths, for example:

λ> [absdir|/home/chris|]</>[absdir|/home/chris|]
<interactive>:23:31-55:
    Couldn't match type ‘Abs’ with ‘Rel’

The equality problem

Paths are now stringently normalized. They have to be a valid path, and they only support single path separators, and all directories are suffixed with a trailing path separator:

λ> $(mkAbsDir "/home/chris//") == $(mkAbsDir "/./home//chris")
True
λ> toFilePath $(mkAbsDir "/home/chris//") ==
   toFilePath $(mkAbsDir "/./home//chris")
True
λ> ($(mkAbsDir "/home/chris//"),toFilePath $(mkAbsDir "/./home//chris"))
("/home/chris/","/home/chris/")

Unpredictable concatenation issues

Because of the stringent normalization, path concatenation, as seen above, is simply string concatenation. This is about as predictable as it can get:

λ> toFilePath $(mkAbsDir "/home/chris//")
"/home/chris/"
λ> toFilePath $(mkRelDir "foo//bar")
"foo/bar/"
λ> [absdir|/home/chris//|]</>[reldir|foo//bar|]
"/home/chris/foo/bar/"

Confusing files and directories

Now that the path type is encoded in the type system, our </> operator prevents improper appending:

λ> [absdir|/home/chris/|]</>[relfile|foo//bar|]
"/home/chris/foo/bar"
λ> [absfile|/home/chris|]</>[relfile|foo//bar|]
<interactive>:35:1-26:
    Couldn't match type ‘File’ with ‘Dir’

Self-documentation

Now I can read the path like:

{ fooPath :: Path Rel Dir, ... }

And know that this refers to the directory relative to some other path, meaning I should be careful to consider the current directory when using this in IO, or that I’ll probably need a parent to append to it at some point.

In practice

We’ve been using this at FP Complete in a number of packages for some months now, it’s turned out surprisingly sufficient for most of our path work with only one bug found. We weren’t sure initially whether it would just be too much of a pain to use, but really it’s quite acceptable given the advantages. You can see its use all over the stack codebase.

Doing I/O

Currently any operations involving I/O can be done by using the existing I/O library:

doesFileExist (toFilePath fp)
readFile (toFilePath fp)

etc. This has problems with respect to accidentally running something like:

doesFileExist $(mkRelDir "foo")

But I/O is currently outside the scope of what this package solves. Once you leave the realm of the Path type invariants are back to your responsibility.

As with the original version of this library, we’re currently building up a set of functions in a Path.IO module over time that fits our real-world use-cases. It may or may not appear in the path package eventually. It’ll need cleaning up and considering what should really be included.

Edit: There is now path-io package that complements the path library and includes complete well-typed interface to directory and temporary. There is work to add more generally useful functions from Stack's Path.IO to it and make Stack depend on the path-io package.

Doing textual manipulations

One problem that crops up sometimes is wanting to manipulate paths. Currently the way we do it is via the filepath library and re-parsing the path:

parseAbsFile . addExtension "/directory/path" "ext" . toFilePath

It doesn’t happen too often, in our experience, to the extent this needs to be more convenient.

Accepting user input

Sometimes you have user input that contains ../. The solution we went with is to have a function like resolveDir (found in path-io package):

resolveDir :: (MonadIO m, MonadThrow m)
           => Path Abs Dir -> FilePath -> m (Path Abs Dir)

Which will call canonicalizePath which collapses and normalizes a path and then we parse with regular old parseAbsDir and we’re cooking with gas. This and others like it might get added to the path package.

Comparing with existing path libraries

filepath and system-filepath

The filepath package is intended as the complimentary package to be used before parsing into a Path value, and/or after printing from a Path value. The package itself contains no type-safety, instead contains a range of cross-platform textual operations. Definitely reach for this library when you want to do more involved manipulations.

The system-filepath package is deprecated in favour of filepath.

system-canonicalpath, canonical-filepath, directory-tree

The system-canonicalpath and the canonical-filepath packages both are a kind of subset of path. They canonicalize a string into an opaque path, but neither distinguish directories from files or absolute/relative. Useful if you just want a canonical path but doesn’t do anything else.

The directory-tree package contains a sum type of dir/file/etc but doesn’t distinguish in its operations relativity or path type.

pathtype

Finally, we come to a path library that path is similar to: the pathtype library. There are the same types of Path Abs File / Path Rel Dir, etc.

The points where this library isn’t enough for me are:

• There is an IsString instance, which means people will use it, and will make mistakes.

• Paths are not normalized into a predictable format, leading to me being unsure when equality will succeed. This is the same problem I encountered in system-filepath. The equality function normalizes, but according to what properties I can reason about? I don’t know.

System.Path.Posix> ("/tmp//" :: Path a Dir) == ("/tmp" :: Path a Dir)
True
System.Path.Posix> ("tmp" :: Path a Dir) == ("/tmp" :: Path a Dir)
True
System.Path.Posix> ("/etc/passwd/" :: Path a b) == ("/etc/passwd" :: Path a b)
True
System.Path.Posix> ("/tmp//" :: Path Abs Dir) == ("/tmp/./" :: Path Abs Dir)
False
System.Path.Posix> ("/tmp/../" :: Path Abs Dir) == ("/" :: Path Abs Dir)
False

• Empty string should not be allowed, and introduction of . due to that gets weird:

System.Path.Posix> fmap getPathString (Right ("." :: Path Rel File))
Right "."
System.Path.Posix> fmap getPathString (mkPathAbsOrRel "")
Right "."
System.Path.Posix> (Right ("." :: Path Rel File)) == (mkPathAbsOrRel "")
False
System.Path.Posix> takeDirectory ("tmp" :: Path Rel Dir)
.
System.Path.Posix> (getPathString ("." :: Path Rel File) ==
                    getPathString ("" :: Path Rel File))
True
System.Path.Posix> (("." :: Path Rel File) == ("" :: Path Rel File))
False

• It has functions like <.>/addExtension which lets you insert an arbitrary string into a path.

• Some functions let you produce nonsense (could be prevented by a stricter type), for example:

System.Path.Posix> takeFileName ("/tmp/" :: Path Abs Dir)
tmp

I’m being a bit picky here, a bit unfair. But the point is really to show the kind of things I tried to avoid in path. In summary, it’s just hard to know where things can go wrong, similar to what was going on in system-filepath.

data-filepath

The data-filepath is also very similar, I discovered it after writing my own at work and was pleased to see it’s mostly the same. The main differences are:

• Uses DataKinds for the relative/absolute and file/dir distinction which as I said above is an overhead.

• Uses a GADT for the path type, which is fine. In my case I wanted to retain the original string which functions that work on the FilePath (String) type already deal with well. It does change the parsing step somewhat, because it parses into segments.

• It’s more lenient at parsing (allowing .. and trailing .).

The API is a bit awkward to just parse a directory, requires a couple functions to get it (going via WeakFilePath), returning only an Either, and there are no functions like parent. But there’s not much to complain about. It’s a fine library, but I didn’t feel the need to drop my own in favor of it. Check it out and decide for yourself.

Summary

There’s a growing interest in making practical use of well-typed file path handling. I think everyone’s wanted it for a while, but few people have really committed to it in practice. Now that I’ve been using path for a while, I can’t really go back. It’ll be interesting to see what new packages crop up in the coming year, I expect there’ll be more.