godirwalk

godirwalk

godirwalk is a library for traversing a directory tree on a file system.

In short, why do I use this library?

1. It's faster than filepath.Walk.
2. It's more correct on Windows than filepath.Walk.
3. It's more easy to use than filepath.Walk.
4. It's more flexible than filepath.Walk.

Usage Example

Additional examples are provided in the examples/ subdirectory.

This library will normalize the provided top level directory name based on the os-specific path separator by calling filepath.Clean on its first argument. However it always provides the pathname created by using the correct os-specific path separator when invoking the provided callback function.

    dirname := "some/directory/root"
    err := godirwalk.Walk(dirname, &godirwalk.Options{
        Callback: func(osPathname string, de *godirwalk.Dirent) error {
            fmt.Printf("%s %s\n", de.ModeType(), osPathname)
            return nil
        },
        Unsorted: true, // (optional) set true for faster yet non-deterministic enumeration (see godoc)
    })

This library not only provides functions for traversing a file system directory tree, but also for obtaining a list of immediate descendants of a particular directory, typically much more quickly than using os.ReadDir or os.ReadDirnames.

Description

Here's why I use godirwalk in preference to filepath.Walk, os.ReadDir, and os.ReadDirnames.

It's faster than filepath.Walk

When compared against filepath.Walk in benchmarks, it has been observed to run between five and ten times the speed on darwin, at speeds comparable to the that of the unix find utility; about twice the speed on linux; and about four times the speed on Windows.

How does it obtain this performance boost? It does less work to give you nearly the same output. This library calls the same syscall functions to do the work, but it makes fewer calls, does not throw away information that it might need, and creates less memory churn along the way by reusing the same scratch buffer for reading from a directory rather than reallocating a new buffer every time it reads file system entry data from the operating system.

While traversing a file system directory tree, filepath.Walk obtains the list of immediate descendants of a directory, and throws away the file system node type information provided by the operating system that comes with the node's name. Then, immediately prior to invoking the callback function, filepath.Walk invokes os.Stat for each node, and passes the returned os.FileInfo information to the callback.

While the os.FileInfo information provided by os.Stat is extremely helpful--and even includes the os.FileMode data--providing it requires an additional system call for each node.

Because most callbacks only care about what the node type is, this library does not throw the type information away, but rather provides that information to the callback function in the form of a os.FileMode value. Note that the provided os.FileMode value that this library provides only has the node type information, and does not have the permission bits, sticky bits, or other information from the file's mode. If the callback does care about a particular node's entire os.FileInfo data structure, the callback can easiy invoke os.Stat when needed, and only when needed.

Benchmarks

$ go test -bench=. -benchmem
goos: darwin
goarch: amd64
pkg: github.com/karrick/godirwalk
BenchmarkReadDirnamesStandardLibrary-12   50000       26250  ns/op       10360  B/op       16  allocs/op
BenchmarkReadDirnamesThisLibrary-12       50000       24372  ns/op        5064  B/op       20  allocs/op
BenchmarkFilepathWalk-12                      1  1099524875  ns/op   228415912  B/op   416952  allocs/op
BenchmarkGodirwalk-12                         2   526754589  ns/op   103110464  B/op   451442  allocs/op
BenchmarkGodirwalkUnsorted-12                 3   509219296  ns/op   100751400  B/op   378800  allocs/op
BenchmarkFlameGraphFilepathWalk-12            1  7478618820  ns/op  2284138176  B/op  4169453  allocs/op
BenchmarkFlameGraphGodirwalk-12               1  4977264058  ns/op  1031105328  B/op  4514423  allocs/op
PASS
ok  	github.com/karrick/godirwalk	21.219s

$ go test -bench=. -benchmem
goos: linux
goarch: amd64
pkg: github.com/karrick/godirwalk
BenchmarkReadDirnamesStandardLibrary-12  100000       15458  ns/op       10360  B/op       16  allocs/op
BenchmarkReadDirnamesThisLibrary-12      100000       14646  ns/op        5064  B/op       20  allocs/op
BenchmarkFilepathWalk-12                      2   631034745  ns/op   228210216  B/op   416939  allocs/op
BenchmarkGodirwalk-12                         3   358714883  ns/op   102988664  B/op   451437  allocs/op
BenchmarkGodirwalkUnsorted-12                 3   355363915  ns/op   100629234  B/op   378796  allocs/op
BenchmarkFlameGraphFilepathWalk-12            1  6086913991  ns/op  2282104720  B/op  4169417  allocs/op
BenchmarkFlameGraphGodirwalk-12               1  3456398824  ns/op  1029886400  B/op  4514373  allocs/op
PASS
ok      github.com/karrick/godirwalk    19.179s

It's more correct on Windows than filepath.Walk

I did not previously care about this either, but humor me. We all love how we can write once and run everywhere. It is essential for the language's adoption, growth, and success, that the software we create can run unmodified on all architectures and operating systems supported by Go.

When the traversed file system has a logical loop caused by symbolic links to directories, on unix filepath.Walk ignores symbolic links and traverses the entire directory tree without error. On Windows however, filepath.Walk will continue following directory symbolic links, even though it is not supposed to, eventually causing filepath.Walk to terminate early and return an error when the pathname gets too long from concatenating endless loops of symbolic links onto the pathname. This error comes from Windows, passes through filepath.Walk, and to the upstream client running filepath.Walk.

The takeaway is that behavior is different based on which platform filepath.Walk is running. While this is clearly not intentional, until it is fixed in the standard library, it presents a compatibility problem.

This library correctly identifies symbolic links that point to directories and will only follow them when FollowSymbolicLinks is set to true. Behavior on Windows and other operating systems is identical.

It's more easy to use than filepath.Walk

Since this library does not invoke os.Stat on every file system node it encounters, there is no possible error event for the callback function to filter on. The third argument in the filepath.WalkFunc function signature to pass the error from os.Stat to the callback function is no longer necessary, and thus eliminated from signature of the callback function from this library.

Also, filepath.Walk invokes the callback function with a solidus delimited pathname regardless of the os-specific path separator. This library invokes the callback function with the os-specific pathname separator, obviating a call to filepath.Clean in the callback function for each node prior to actually using the provided pathname.

In other words, even on Windows, filepath.Walk will invoke the callback with some/path/to/foo.txt, requiring well written clients to perform pathname normalization for every file prior to working with the specified file. In truth, many clients developed on unix and not tested on Windows neglect this subtlety, and will result in software bugs when running on Windows. This library would invoke the callback function with some\path\to\foo.txt for the same file when running on Windows, eliminating the need to normalize the pathname by the client, and lessen the likelyhood that a client will work on unix but not on Windows.

It's more flexible than filepath.Walk

Configurable Handling of Symbolic Links

The default behavior of this library is to ignore symbolic links to directories when walking a directory tree, just like filepath.Walk does. However, it does invoke the callback function with each node it finds, including symbolic links. If a particular use case exists to follow symbolic links when traversing a directory tree, this library can be invoked in manner to do so, by setting the FollowSymbolicLinks parameter to true.

Configurable Sorting of Directory Children

The default behavior of this library is to always sort the immediate descendants of a directory prior to visiting each node, just like filepath.Walk does. This is usually the desired behavior. However, this does come at slight performance and memory penalties required to sort the names when a directory node has many entries. Additionally if caller specifies Unsorted enumeration, reading directories is lazily performed as the caller consumes entries. If a particular use case exists that does not require sorting the directory's immediate descendants prior to visiting its nodes, this library will skip the sorting step when the Unsorted parameter is set to true.

Configurable Post Children Callback

This library provides upstream code with the ability to specify a callback to be invoked for each directory after its children are processed. This has been used to recursively delete empty directories after traversing the file system in a more efficient manner. See the examples/clean-empties directory for an example of this usage.

Configurable Error Callback

This library provides upstream code with the ability to specify a callback to be invoked for errors that the operating system returns, allowing the upstream code to determine the next course of action to take, whether to halt walking the hierarchy, as it would do were no error callback provided, or skip the node that caused the error. See the examples/walk-fast directory for an example of this usage.