|Title:||os.scandir() function -- a better and faster directory iterator|
|Author:||Ben Hoyt <benhoyt at gmail.com>|
|BDFL-Delegate:||Victor Stinner <vstinner at python.org>|
|Post-History:||27-Jun-2014, 8-Jul-2014, 14-Jul-2014|
- Specifics of proposal
- Use in the wild
- Rejected ideas
- Wildcard support
- Methods not following symlinks by default
- DirEntry attributes being properties
- DirEntry fields being "static" attribute-only objects
- DirEntry fields being static with an ensure_lstat option
- Return values being (name, stat_result) two-tuples
- Return values being overloaded stat_result objects
- Return values being pathlib.Path objects
- Possible improvements
- Previous discussion
This PEP proposes including a new directory iteration function, os.scandir(), in the standard library. This new function adds useful functionality and increases the speed of os.walk() by 2-20 times (depending on the platform and file system) by avoiding calls to os.stat() in most cases.
Python's built-in os.walk() is significantly slower than it needs to be, because -- in addition to calling os.listdir() on each directory -- it executes the stat() system call or GetFileAttributes() on each file to determine whether the entry is a directory or not.
But the underlying system calls -- FindFirstFile / FindNextFile on Windows and readdir on POSIX systems -- already tell you whether the files returned are directories or not, so no further system calls are needed. Further, the Windows system calls return all the information for a stat_result object on the directory entry, such as file size and last modification time.
In short, you can reduce the number of system calls required for a tree function like os.walk() from approximately 2N to N, where N is the total number of files and directories in the tree. (And because directory trees are usually wider than they are deep, it's often much better than this.)
In practice, removing all those extra system calls makes os.walk() about 8-9 times as fast on Windows, and about 2-3 times as fast on POSIX systems. So we're not talking about micro- optimizations. See more benchmarks here.
Somewhat relatedly, many people (see Python Issue 11406) are also keen on a version of os.listdir() that yields filenames as it iterates instead of returning them as one big list. This improves memory efficiency for iterating very large directories.
So, as well as providing a scandir() iterator function for calling directly, Python's existing os.walk() function can be sped up a huge amount.
The implementation of this proposal was written by Ben Hoyt (initial version) and Tim Golden (who helped a lot with the C extension module). It lives on GitHub at benhoyt/scandir. (The implementation may lag behind the updates to this PEP a little.)
Note that this module has been used and tested (see "Use in the wild" section in this PEP), so it's more than a proof-of-concept. However, it is marked as beta software and is not extensively battle-tested. It will need some cleanup and more thorough testing before going into the standard library, as well as integration into posixmodule.c.
Specifically, this PEP proposes adding a single function to the os module in the standard library, scandir, that takes a single, optional string as its argument:
scandir(path='.') -> generator of DirEntry objects
Like listdir, scandir calls the operating system's directory iteration system calls to get the names of the files in the given path, but it's different from listdir in two ways:
- Instead of returning bare filename strings, it returns lightweight DirEntry objects that hold the filename string and provide simple methods that allow access to the additional data the operating system may have returned.
- It returns a generator instead of a list, so that scandir acts as a true iterator instead of returning the full list immediately.
scandir() yields a DirEntry object for each file and sub-directory in path. Just like listdir, the '.' and '..' pseudo-directories are skipped, and the entries are yielded in system-dependent order. Each DirEntry object has the following attributes and methods:
- name: the entry's filename, relative to the scandir path argument (corresponds to the return values of os.listdir)
- path: the entry's full path name (not necessarily an absolute path) -- the equivalent of os.path.join(scandir_path, entry.name)
- inode(): return the inode number of the entry. The result is cached on the DirEntry object, use os.stat(entry.path, follow_symlinks=False).st_ino to fetch up-to-date information. On Unix, no system call is required.
- is_dir(*, follow_symlinks=True): similar to pathlib.Path.is_dir(), but the return value is cached on the DirEntry object; doesn't require a system call in most cases; don't follow symbolic links if follow_symlinks is False
- is_file(*, follow_symlinks=True): similar to pathlib.Path.is_file(), but the return value is cached on the DirEntry object; doesn't require a system call in most cases; don't follow symbolic links if follow_symlinks is False
- is_symlink(): similar to pathlib.Path.is_symlink(), but the return value is cached on the DirEntry object; doesn't require a system call in most cases
- stat(*, follow_symlinks=True): like os.stat(), but the return value is cached on the DirEntry object; does not require a system call on Windows (except for symlinks); don't follow symbolic links (like os.lstat()) if follow_symlinks is False
All methods may perform system calls in some cases and therefore possibly raise OSError -- see the "Notes on exception handling" section for more details.
The DirEntry attribute and method names were chosen to be the same as those in the new pathlib module where possible, for consistency. The only difference in functionality is that the DirEntry methods cache their values on the entry object after the first call.
Like the other functions in the os module, scandir() accepts either a bytes or str object for the path parameter, and returns the DirEntry.name and DirEntry.path attributes with the same type as path. However, it is strongly recommended to use the str type, as this ensures cross-platform support for Unicode filenames. (On Windows, bytes filenames have been deprecated since Python 3.3).
As part of this proposal, os.walk() will also be modified to use scandir() rather than listdir() and os.path.isdir(). This will increase the speed of os.walk() very significantly (as mentioned above, by 2-20 times, depending on the system).
First, a very simple example of scandir() showing use of the DirEntry.name attribute and the DirEntry.is_dir() method:
def subdirs(path): """Yield directory names not starting with '.' under given path.""" for entry in os.scandir(path): if not entry.name.startswith('.') and entry.is_dir(): yield entry.name
This subdirs() function will be significantly faster with scandir than os.listdir() and os.path.isdir() on both Windows and POSIX systems, especially on medium-sized or large directories.
Or, for getting the total size of files in a directory tree, showing use of the DirEntry.stat() method and DirEntry.path attribute:
def get_tree_size(path): """Return total size of files in given path and subdirs.""" total = 0 for entry in os.scandir(path): if entry.is_dir(follow_symlinks=False): total += get_tree_size(entry.path) else: total += entry.stat(follow_symlinks=False).st_size return total
This also shows the use of the follow_symlinks parameter to is_dir() -- in a recursive function like this, we probably don't want to follow links. (To properly follow links in a recursive function like this we'd want special handling for the case where following a symlink leads to a recursive loop.)
Note that get_tree_size() will get a huge speed boost on Windows, because no extra stat call are needed, but on POSIX systems the size information is not returned by the directory iteration functions, so this function won't gain anything there.
The DirEntry objects are relatively dumb -- the name and path attributes are obviously always cached, and the is_X and stat methods cache their values (immediately on Windows via FindNextFile, and on first use on POSIX systems via a stat system call) and never refetch from the system.
For this reason, DirEntry objects are intended to be used and thrown away after iteration, not stored in long-lived data structured and the methods called again and again.
If developers want "refresh" behaviour (for example, for watching a file's size change), they can simply use pathlib.Path objects, or call the regular os.stat() or os.path.getsize() functions which get fresh data from the operating system every call.
DirEntry.is_X() and DirEntry.stat() are explicitly methods rather than attributes or properties, to make it clear that they may not be cheap operations (although they often are), and they may do a system call. As a result, these methods may raise OSError.
For example, DirEntry.stat() will always make a system call on POSIX-based systems, and the DirEntry.is_X() methods will make a stat() system call on such systems if readdir() does not support d_type or returns a d_type with a value of DT_UNKNOWN, which can occur under certain conditions or on certain file systems.
Often this does not matter -- for example, os.walk() as defined in the standard library only catches errors around the listdir() calls.
Also, because the exception-raising behaviour of the DirEntry.is_X methods matches that of pathlib -- which only raises OSError in the case of permissions or other fatal errors, but returns False if the path doesn't exist or is a broken symlink -- it's often not necessary to catch errors around the is_X() calls.
However, when a user requires fine-grained error handling, it may be desirable to catch OSError around all method calls and handle as appropriate.
For example, below is a version of the get_tree_size() example shown above, but with fine-grained error handling added:
def get_tree_size(path): """Return total size of files in path and subdirs. If is_dir() or stat() fails, print an error message to stderr and assume zero size (for example, file has been deleted). """ total = 0 for entry in os.scandir(path): try: is_dir = entry.is_dir(follow_symlinks=False) except OSError as error: print('Error calling is_dir():', error, file=sys.stderr) continue if is_dir: total += get_tree_size(entry.path) else: try: total += entry.stat(follow_symlinks=False).st_size except OSError as error: print('Error calling stat():', error, file=sys.stderr) return total
The scandir module on GitHub has been forked and used quite a bit (see "Use in the wild" in this PEP), but there's also been a fair bit of direct support for a scandir-like function from core developers and others on the python-dev and python-ideas mailing lists. A sampling:
- python-dev: a good number of +1's and very few negatives for scandir and PEP 471 on this June 2014 python-dev thread
- Nick Coghlan, a core Python developer: "I've had the local Red Hat release engineering team express their displeasure at having to stat every file in a network mounted directory tree for info that is present in the dirent structure, so a definite +1 to os.scandir from me, so long as it makes that info available." [source1]
- Tim Golden, a core Python developer, supports scandir enough to have spent time refactoring and significantly improving scandir's C extension module. [source2]
- Christian Heimes, a core Python developer: "+1 for something like yielddir()" [source3] and "Indeed! I'd like to see the feature in 3.4 so I can remove my own hack from our code base." [source4]
- Gregory P. Smith, a core Python developer: "As 3.4beta1 happens tonight, this isn't going to make 3.4 so i'm bumping this to 3.5. I really like the proposed design outlined above." [source5]
- Guido van Rossum on the possibility of adding scandir to Python 3.5 (as it was too late for 3.4): "The ship has likewise sailed for adding scandir() (whether to os or pathlib). By all means experiment and get it ready for consideration for 3.5, but I don't want to add it to 3.4." [source6]
Support for this PEP itself (meta-support?) was given by Nick Coghlan on python-dev: "A PEP reviewing all this for 3.5 and proposing a specific os.scandir API would be a good thing." [source7]
To date, the scandir implementation is definitely useful, but has been clearly marked "beta", so it's uncertain how much use of it there is in the wild. Ben Hoyt has had several reports from people using it. For example:
- Chris F: "I am processing some pretty large directories and was half expecting to have to modify getdents. So thanks for saving me the effort." [via personal email]
- bschollnick: "I wanted to let you know about this, since I am using Scandir as a building block for this code. Here's a good example of scandir making a radical performance improvement over os.listdir." [source8]
- Avram L: "I'm testing our scandir for a project I'm working on. Seems pretty solid, so first thing, just want to say nice work!" [via personal email]
- Matt Z: "I used scandir to dump the contents of a network dir in under 15 seconds. 13 root dirs, 60,000 files in the structure. This will replace some old VBA code embedded in a spreadsheet that was taking 15-20 minutes to do the exact same thing." [via personal email]
GitHub stats don't mean too much, but scandir does have several watchers, issues, forks, etc. Here's the run-down as of the stats as of July 7, 2014:
- Watchers: 17
- Stars: 57
- Forks: 20
- Issues: 4 open, 26 closed
Also, because this PEP will increase the speed of os.walk() significantly, there are thousands of developers and scripts, and a lot of production code, that would benefit from it. For example, on GitHub, there are almost as many uses of os.walk (194,000) as there are of os.mkdir (230,000).
The only other real contender for this function's name was iterdir(). However, iterX() functions in Python (mostly found in Python 2) tend to be simple iterator equivalents of their non-iterator counterparts. For example, dict.iterkeys() is just an iterator version of dict.keys(), but the objects returned are identical. In scandir()'s case, however, the return values are quite different objects (DirEntry objects vs filename strings), so this should probably be reflected by a difference in name -- hence scandir().
See some relevant discussion on python-dev.
FindFirstFile/FindNextFile on Windows support passing a "wildcard" like *.jpg, so at first folks (this PEP's author included) felt it would be a good idea to include a windows_wildcard keyword argument to the scandir function so users could pass this in.
However, on further thought and discussion it was decided that this would be bad idea, unless it could be made cross-platform (a pattern keyword argument or similar). This seems easy enough at first -- just use the OS wildcard support on Windows, and something like fnmatch or re afterwards on POSIX-based systems.
Unfortunately the exact Windows wildcard matching rules aren't really documented anywhere by Microsoft, and they're quite quirky (see this blog post), meaning it's very problematic to emulate using fnmatch or regexes.
So the consensus was that Windows wildcard support was a bad idea. It would be possible to add at a later date if there's a cross-platform way to achieve it, but not for the initial version.
In some ways it would be nicer for the DirEntry is_X() and stat() to be properties instead of methods, to indicate they're very cheap or free. However, this isn't quite the case, as stat() will require an OS call on POSIX-based systems but not on Windows. Even is_dir() and friends may perform an OS call on POSIX-based systems if the dirent.d_type value is DT_UNKNOWN (on certain file systems).
Also, people would expect the attribute access entry.is_dir to only ever raise AttributeError, not OSError in the case it makes a system call under the covers. Calling code would have to have a try/except around what looks like a simple attribute access, and so it's much better to make them methods.
See this May 2013 python-dev thread where this PEP author makes this case and there's agreement from a core developers.
In this July 2014 python-dev message, Paul Moore suggested a solution that was a "thin wrapper round the OS feature", where the DirEntry object had only static attributes: name, path, and is_X, with the st_X attributes only present on Windows. The idea was to use this simpler, lower-level function as a building block for higher-level functions.
At first there was general agreement that simplifying in this way was a good thing. However, there were two problems with this approach. First, the assumption is the is_dir and similar attributes are always present on POSIX, which isn't the case (if d_type is not present or is DT_UNKNOWN). Second, it's a much harder-to-use API in practice, as even the is_dir attributes aren't always present on POSIX, and would need to be tested with hasattr() and then os.stat() called if they weren't present.
See this July 2014 python-dev response from this PEP's author detailing why this option is a non-ideal solution, and the subsequent reply from Paul Moore voicing agreement.
Another seemingly simpler and attractive option was suggested by Nick Coghlan in this June 2014 python-dev message: make DirEntry.is_X and DirEntry.lstat_result properties, and populate DirEntry.lstat_result at iteration time, but only if the new argument ensure_lstat=True was specified on the scandir() call.
This does have the advantage over the above in that you can easily get the stat result from scandir() if you need it. However, it has the serious disadvantage that fine-grained error handling is messy, because stat() will be called (and hence potentially raise OSError) during iteration, leading to a rather ugly, hand-made iteration loop:
it = os.scandir(path) while True: try: entry = next(it) except OSError as error: handle_error(path, error) except StopIteration: break
Or it means that scandir() would have to accept an onerror argument -- a function to call when stat() errors occur during iteration. This seems to this PEP's author neither as direct nor as Pythonic as try/except around a DirEntry.stat() call.
Another drawback is that os.scandir() is written to make code faster. Always calling os.lstat() on POSIX would not bring any speedup. In most cases, you don't need the full stat_result object -- the is_X() methods are enough and this information is already known.
Initially this PEP's author proposed this concept as a function called iterdir_stat() which yielded two-tuples of (name, stat_result). This does have the advantage that there are no new types introduced. However, the stat_result is only partially filled on POSIX-based systems (most fields set to None and other quirks), so they're not really stat_result objects at all, and this would have to be thoroughly documented as different from os.stat().
Also, Python has good support for proper objects with attributes and methods, which makes for a saner and simpler API than two-tuples. It also makes the DirEntry objects more extensible and future-proof as operating systems add functionality and we want to include this in DirEntry.
See also some previous discussion:
Another alternative discussed was making the return values to be overloaded stat_result objects with name and path attributes. However, apart from this being a strange (and strained!) kind of overloading, this has the same problems mentioned above -- most of the stat_result information is not fetched by readdir() on POSIX systems, only (part of) the st_mode value.
With Antoine Pitrou's new standard library pathlib module, it at first seems like a great idea for scandir() to return instances of pathlib.Path. However, pathlib.Path's is_X() and stat() functions are explicitly not cached, whereas scandir has to cache them by design, because it's (often) returning values from the original directory iteration system call.
And if the pathlib.Path instances returned by scandir cached stat values, but the ordinary pathlib.Path objects explicitly don't, that would be more than a little confusing.
Guido van Rossum explicitly rejected pathlib.Path caching stat in the context of scandir here, making pathlib.Path objects a bad choice for scandir return values.
There are many possible improvements one could make to scandir, but here is a short list of some this PEP's author has in mind:
- scandir could potentially be further sped up by calling readdir / FindNextFile say 50 times per Py_BEGIN_ALLOW_THREADS block so that it stays in the C extension module for longer, and may be somewhat faster as a result. This approach hasn't been tested, but was suggested by on Issue 11406 by Antoine Pitrou. [source9]
- scandir could use a free list to avoid the cost of memory allocation for each iteration -- a short free list of 10 or maybe even 1 may help. Suggested by Victor Stinner on a python-dev thread on June 27.
- Original November 2012 thread Ben Hoyt started on python-ideas about speeding up os.walk()
- Python Issue 11406, which includes the original proposal for a scandir-like function
- Further May 2013 thread Ben Hoyt started on python-dev that refined the scandir() API, including Nick Coghlan's suggestion of scandir yielding DirEntry-like objects
- November 2013 thread Ben Hoyt started on python-dev to discuss the interaction between scandir and the new pathlib module
- June 2014 thread Ben Hoyt started on python-dev to discuss the first version of this PEP, with extensive discussion about the API
- First July 2014 thread Ben Hoyt started on python-dev to discuss his updates to PEP 471
- Second July 2014 thread Ben Hoyt started on python-dev to discuss the remaining decisions needed to finalize PEP 471, specifically whether the DirEntry methods should follow symlinks by default
- Question on StackOverflow about why os.walk() is slow and pointers on how to fix it (this inspired the author of this PEP early on)
- BetterWalk, this PEP's author's previous attempt at this, on which the scandir code is based
This document has been placed in the public domain.