class org.apache.hadoop.fs.FileSystem

The abstract FileSystem class is the original class to access Hadoop filesystems; non-abstract subclasses exist for all Hadoop-supported filesystems.

All operations that take a Path to this interface MUST support relative paths. In such a case, they must be resolved relative to the working directory defined by setWorkingDirectory().

For all clients, therefore, we also add the notion of a state component PWD: this represents the present working directory of the client. Changes to this state are not reflected in the filesystem itself: they are unique to the instance of the client.

Implementation Note: the static FileSystem get(URI uri, Configuration conf) method MAY return a pre-existing instance of a filesystem client class—a class that may also be in use in other threads. The implementations of FileSystem shipped with Apache Hadoop do not make any attempt to synchronize access to the working directory field.


All the requirements of a valid FileSystem are considered implicit preconditions and postconditions: all operations on a valid FileSystem MUST result in a new FileSystem that is also valid.

Feasible features

Protected directories

HDFS has the notion of Protected Directories, which are declared in the option fs.protected.directories. Any attempt to delete or rename such a directory or a parent thereof raises an AccessControlException. Accordingly, any attempt to delete the root directory SHALL, if there is a protected directory, result in such an exception being raised.

Predicates and other state access operations

boolean exists(Path p)

def exists(FS, p) = p in paths(FS)

boolean isDirectory(Path p)

def isDir(FS, p) = p in directories(FS)

boolean isFile(Path p)

def isFile(FS, p) = p in filenames(FS)

FileStatus getFileStatus(Path p)

Get the status of a path


if not exists(FS, p) : raise FileNotFoundException


result = stat: FileStatus where:
    if isFile(FS, p) :
        stat.length = len(FS.Files[p])
        stat.isdir = False
        stat.blockSize > 0
    elif isDir(FS, p) :
        stat.length = 0
        stat.isdir = True
    elif isSymlink(FS, p) :
        stat.length = 0
        stat.isdir = False
        stat.symlink = FS.Symlinks[p]
    stat.hasAcl = hasACL(FS, p)
    stat.isEncrypted = inEncryptionZone(FS, p)
    stat.isErasureCoded = isErasureCoded(FS, p)

The returned FileStatus status of the path additionally carries details on ACL, encryption and erasure coding information. getFileStatus(Path p).hasAcl() can be queried to find if the path has an ACL. getFileStatus(Path p).isEncrypted() can be queried to find if the path is encrypted. getFileStatus(Path p).isErasureCoded() will tell if the path is erasure coded or not.

YARN’s distributed cache lets applications add paths to be cached across containers and applications via Job.addCacheFile() and Job.addCacheArchive(). The cache treats world-readable resources paths added as shareable across applications, and downloads them differently, unless they are declared as encrypted.

To avoid failures during container launching, especially when delegation tokens are used, filesystems and object stores which not implement POSIX access permissions for both files and directories, MUST always return true to the isEncrypted() predicate. This can be done by setting the encrypted flag to true when creating the FileStatus instance.


Synchronize metadata state of the client with the latest state of the metadata service of the FileSystem.

In highly available FileSystems standby service can be used as a read-only metadata replica. This call is essential to guarantee consistency of reads from the standby replica and to avoid stale reads.

It is currently only implemented for HDFS and others will just throw UnsupportedOperationException.



This call internally records the state of the metadata service at the time of the call. This guarantees consistency of subsequent reads from any metadata replica. It assures the client will never access the state of the metadata that preceded the recorded state.

HDFS implementation notes

HDFS supports msync() in HA mode by calling the Active NameNode and requesting its latest journal transaction ID. For more details see HDFS documentation Consistent Reads from HDFS Observer NameNode

Path getHomeDirectory()

The function getHomeDirectory returns the home directory for the FileSystem and the current user account.

For some FileSystems, the path is ["/", "users", System.getProperty("user-name")].

However, for HDFS, the username is derived from the credentials used to authenticate the client with HDFS. This may differ from the local user account name.

It is the responsibility of the FileSystem to determine the actual home directory of the caller.



result = p where valid-path(FS, p)

There is no requirement that the path exists at the time the method was called, or, if it exists, that it points to a directory. However, code tends to assume that not isFile(FS, getHomeDirectory()) holds to the extent that follow-on code may fail.

Implementation Notes

  • The FTPFileSystem queries this value from the remote filesystem and may fail with a RuntimeException or subclass thereof if there is a connectivity problem. The time to execute the operation is not bounded.

FileStatus[] listStatus(Path path, PathFilter filter)

Lists entries under a path, path.

If path refers to a file and the filter accepts it, then that file’s FileStatus entry is returned in a single-element array.

If the path refers to a directory, the call returns a list of all its immediate child paths which are accepted by the filter —and does not include the directory itself.

A PathFilter filter is a class whose accept(path) returns true iff the path path meets the filter’s conditions.


Path path must exist:

if not exists(FS, path) : raise FileNotFoundException


if isFile(FS, path) and filter.accept(path) :
  result = [ getFileStatus(path) ]

elif isFile(FS, path) and not filter.accept(P) :
  result = []

elif isDir(FS, path):
  result = [
    getFileStatus(c) for c in children(FS, path) if filter.accepts(c)

Implicit invariant: the contents of a FileStatus of a child retrieved via listStatus() are equal to those from a call of getFileStatus() to the same path:

forall fs in listStatus(path) :
  fs == getFileStatus(fs.path)

Ordering of results: there is no guarantee of ordering of the listed entries. While HDFS currently returns an alphanumerically sorted list, neither the Posix readdir() nor Java’s File.listFiles() API calls define any ordering of returned values. Applications which require a uniform sort order on the results must perform the sorting themselves.

Null return: Local filesystems prior to 3.0.0 returned null upon access error. It is considered erroneous. Expect IOException upon access error.

Atomicity and Consistency

By the time the listStatus() operation returns to the caller, there is no guarantee that the information contained in the response is current. The details MAY be out of date, including the contents of any directory, the attributes of any files, and the existence of the path supplied.

The state of a directory MAY change during the evaluation process.

  • After an entry at path P is created, and before any other changes are made to the filesystem, listStatus(P) MUST find the file and return its status.

  • After an entry at path P is deleted, and before any other changes are made to the filesystem, listStatus(P) MUST raise a FileNotFoundException.

  • After an entry at path P is created, and before any other changes are made to the filesystem, the result of listStatus(parent(P)) SHOULD include the value of getFileStatus(P).

  • After an entry at path P is deleted, and before any other changes are made to the filesystem, the result of listStatus(parent(P)) SHOULD NOT include the value of getFileStatus(P).

This is not a theoretical possibility, it is observable in HDFS when a directory contains many thousands of files.

Consider a directory "/d" with the contents:


If the number of files is such that HDFS returns a partial listing in each response, then, if a listing listStatus("/d") takes place concurrently with the operation rename("/d/a","/d/z")), the result may be one of:

[a, part-0000001, ... , part-9999999]
[part-0000001, ... , part-9999999, z]
[a, part-0000001, ... , part-9999999, z]
[part-0000001, ... , part-9999999]

While this situation is likely to be a rare occurrence, it MAY happen. In HDFS these inconsistent views are only likely when listing a directory with many children.

Other filesystems may have stronger consistency guarantees, or return inconsistent data more readily.

FileStatus[] listStatus(Path path)

This is exactly equivalent to listStatus(Path, DEFAULT_FILTER) where DEFAULT_FILTER.accept(path) = True for all paths.

The atomicity and consistency constraints are as for listStatus(Path, DEFAULT_FILTER).

FileStatus[] listStatus(Path[] paths, PathFilter filter)

Enumerate all files found in the list of directories passed in, calling listStatus(path, filter) on each one.

As with listStatus(path, filter), the results may be inconsistent. That is: the state of the filesystem changed during the operation.

There are no guarantees as to whether paths are listed in a specific order, only that they must all be listed, and, at the time of listing, exist.


All paths must exist. There is no requirement for uniqueness.

forall p in paths :
  exists(FS, p) else raise FileNotFoundException


The result is an array whose entries contain every status element found in the path listings, and no others.

result = [listStatus(p, filter) for p in paths]

Implementations MAY merge duplicate entries; and/or optimize the operation by recoginizing duplicate paths and only listing the entries once.

The default implementation iterates through the list; it does not perform any optimizations.

The atomicity and consistency constraints are as for listStatus(Path, PathFilter).

RemoteIterator<FileStatus> listStatusIterator(Path p)

Return an iterator enumerating the FileStatus entries under a path. This is similar to listStatus(Path) except the fact that rather than returning an entire list, an iterator is returned. The result is exactly the same as listStatus(Path), provided no other caller updates the directory during the listing. Having said that, this does not guarantee atomicity if other callers are adding/deleting the files inside the directory while listing is being performed. Different filesystems may provide a more efficient implementation, for example S3A does the listing in pages and fetches the next pages asynchronously while a page is getting processed.

Note that now since the initial listing is async, bucket/path existence exception may show up later during next() call.

Callers should prefer using listStatusIterator over listStatus as it is incremental in nature.

FileStatus[] listStatus(Path[] paths)

Enumerate all files found in the list of directories passed in, calling listStatus(path, DEFAULT_FILTER) on each one, where the DEFAULT_FILTER accepts all path names.

RemoteIterator[LocatedFileStatus] listLocatedStatus(Path path, PathFilter filter)

Return an iterator enumerating the LocatedFileStatus entries under a path. This is similar to listStatus(Path) except that the return value is an instance of the LocatedFileStatus subclass of a FileStatus, and that rather than return an entire list, an iterator is returned.

This is actually a protected method, directly invoked by listLocatedStatus(Path path). Calls to it may be delegated through layered filesystems, such as FilterFileSystem, so its implementation MUST be considered mandatory, even if listLocatedStatus(Path path) has been implemented in a different manner. There are open JIRAs proposing making this method public; it may happen in future.

There is no requirement for the iterator to provide a consistent view of the child entries of a path. The default implementation does use listStatus(Path) to list its children, with its consistency constraints already documented. Other implementations may perform the enumeration even more dynamically. For example fetching a windowed subset of child entries, so avoiding building up large data structures and the transmission of large messages. In such situations, changes to the filesystem are more likely to become visible.

Callers MUST assume that the iteration operation MAY fail if changes to the filesystem take place between this call returning and the iteration being completely performed.


Path path must exist:

if not exists(FS, path) : raise FileNotFoundException


The operation generates a set of results, resultset, equal to the result of listStatus(path, filter):

if isFile(FS, path) and filter.accept(path) :
  resultset =  [ getLocatedFileStatus(FS, path) ]

elif isFile(FS, path) and not filter.accept(path) :
  resultset = []

elif isDir(FS, path) :
  resultset = [
    getLocatedFileStatus(FS, c)
     for c in children(FS, path) where filter.accept(c)

The operation getLocatedFileStatus(FS, path: Path): LocatedFileStatus is defined as a generator of a LocatedFileStatus instance ls where:

fileStatus = getFileStatus(FS, path)

bl = getFileBlockLocations(FS, path, 0, fileStatus.len)

locatedFileStatus = new LocatedFileStatus(fileStatus, bl)

The ordering in which the elements of resultset are returned in the iterator is undefined.

The atomicity and consistency constraints are as for listStatus(Path, PathFilter).

RemoteIterator[LocatedFileStatus] listLocatedStatus(Path path)

The equivalent to listLocatedStatus(path, DEFAULT_FILTER), where DEFAULT_FILTER accepts all path names.

RemoteIterator[LocatedFileStatus] listFiles(Path path, boolean recursive)

Create an iterator over all files in/under a directory, potentially recursing into child directories.

The goal of this operation is to permit large recursive directory scans to be handled more efficiently by filesystems, by reducing the amount of data which must be collected in a single RPC call.


if not exists(FS, path) : raise FileNotFoundException


The outcome is an iterator, whose output from the sequence of calls can be defined as the set iteratorset:

if not recursive:
  iteratorset == listStatus(path)
  iteratorset = [
    getLocatedFileStatus(FS, d)
      for d in descendants(FS, path)

The function getLocatedFileStatus(FS, d) is as defined in listLocatedStatus(Path, PathFilter).

The atomicity and consistency constraints are as for listStatus(Path, PathFilter).

ContentSummary getContentSummary(Path path)

Given a path return its content summary.

getContentSummary() first checks if the given path is a file and if yes, it returns 0 for directory count and 1 for file count.


if not exists(FS, path) : raise FileNotFoundException


Returns a ContentSummary object with information such as directory count and file count for a given path.

The atomicity and consistency constraints are as for listStatus(Path, PathFilter).

BlockLocation[] getFileBlockLocations(FileStatus f, int s, int l)


if s < 0 or l < 0 : raise {HadoopIllegalArgumentException, InvalidArgumentException}
  • HDFS throws HadoopIllegalArgumentException for an invalid offset or length; this extends IllegalArgumentException.


If the filesystem is location aware, it must return the list of block locations where the data in the range [s:s+l] can be found.

if f == null :
    result = null
elif f.getLen() <= s:
    result = []
else result = [ locations(FS, b) for b in blocks(FS, p, s, s+l)]


  def locations(FS, b) = a list of all locations of a block in the filesystem

  def blocks(FS, p, s, s +  l)  = a list of the blocks containing data(FS, path)[s:s+l]

Note that as length(FS, f) is defined as 0 if isDir(FS, f), the result of getFileBlockLocations() on a directory is []

If the filesystem is not location aware, it SHOULD return

    BlockLocation(["localhost:9866"] ,
               0, f.getLen())
   ] ;

*A bug in Hadoop 1.0.3 means that a topology path of the same number of elements as the cluster topology MUST be provided, hence Filesystems SHOULD return that "/default/localhost" path. While this is no longer an issue, the convention is generally retained.

BlockLocation[] getFileBlockLocations(Path P, int S, int L)


if p == null : raise NullPointerException
if not exists(FS, p) : raise FileNotFoundException


result = getFileBlockLocations(getFileStatus(FS, P), S, L)

long getDefaultBlockSize()

Get the “default” block size for a filesystem. This is often used during split calculations to divide work optimally across a set of worker processes.



result = integer > 0

Although there is no defined minimum value for this result, as it is used to partition work during job submission, a block size that is too small will result in badly partitioned workload, or even the JobSubmissionClient and equivalent running out of memory as it calculates the partitions.

Any FileSystem that does not actually break files into blocks SHOULD return a number for this that results in efficient processing. A FileSystem MAY make this user-configurable (the object store connectors usually do this).

long getDefaultBlockSize(Path p)

Get the “default” block size for a path –that is, the block size to be used when writing objects to a path in the filesystem.



result = integer >= 0

The outcome of this operation is usually identical to getDefaultBlockSize(), with no checks for the existence of the given path.

Filesystems that support mount points may have different default values for different paths, in which case the specific default value for the destination path SHOULD be returned.

It is not an error if the path does not exist: the default/recommended value for that part of the filesystem MUST be returned.

long getBlockSize(Path p)

This method is exactly equivalent to querying the block size of the FileStatus structure returned in getFileStatus(p). It is deprecated in order to encourage users to make a single call to getFileStatus(p) and then use the result to examine multiple attributes of the file (e.g. length, type, block size). If more than one attribute is queried, This can become a significant performance optimization —and reduce load on the filesystem.


if not exists(FS, p) : raise FileNotFoundException


if len(FS, P) > 0 :  getFileStatus(P).getBlockSize() > 0
result == getFileStatus(P).getBlockSize()
  1. The outcome of this operation MUST be identical to the value of getFileStatus(P).getBlockSize().
  2. By inference, it MUST be > 0 for any file of length > 0.

Path getEnclosingRoot(Path p)

This method is used to find a root directory for a path given. This is useful for creating staging and temp directories in the same enclosing root directory. There are constraints around how renames are allowed to atomically occur (ex. across hdfs volumes or across encryption zones).

For any two paths p1 and p2 that do not have the same enclosing root, rename(p1, p2) is expected to fail or will not be atomic.

For object stores, even with the same enclosing root, there is no guarantee file or directory rename is atomic

The following statement is always true: getEnclosingRoot(p) == getEnclosingRoot(getEnclosingRoot(p))

path in ancestors(FS, p) or path == p:
isDir(FS, p)


The path does not have to exist, but the path does need to be valid and reconcilable by the filesystem * if a linkfallback is used all paths are reconcilable * if a linkfallback is not used there must be a mount point covering the path


  • The path returned will not be null, if there is no deeper enclosing root, the root path (“/”) will be returned.
  • The path returned is a directory

State Changing Operations

boolean mkdirs(Path p, FsPermission permission)

Create a directory and all its parents.


The path must either be a directory or not exist

 if exists(FS, p) and not isDir(FS, p) :
     raise [ParentNotDirectoryException, FileAlreadyExistsException, IOException]

No ancestor may be a file

forall d = ancestors(FS, p) : 
    if exists(FS, d) and not isDir(FS, d) :
        raise {ParentNotDirectoryException, FileAlreadyExistsException, IOException}


FS' where FS'.Directories = FS.Directories + [p] + ancestors(FS, p)
result = True

The condition exclusivity requirement of a FileSystem’s directories, files and symbolic links must hold.

The probe for the existence and type of a path and directory creation MUST be atomic. The combined operation, including mkdirs(parent(F)) MAY be atomic.

The return value is always true—even if a new directory is not created (this is defined in HDFS).

FSDataOutputStream create(Path, ...)

FSDataOutputStream create(Path p,
      FsPermission permission,
      boolean overwrite,
      int bufferSize,
      short replication,
      long blockSize,
      Progressable progress) throws IOException;


The file must not exist for a no-overwrite create:

if not overwrite and isFile(FS, p) : raise FileAlreadyExistsException

Writing to or overwriting a directory must fail.

if isDir(FS, p) : raise {FileAlreadyExistsException, FileNotFoundException, IOException}

No ancestor may be a file

forall d = ancestors(FS, p) : 
    if exists(FS, d) and not isDir(FS, d) :
        raise {ParentNotDirectoryException, FileAlreadyExistsException, IOException}

FileSystems may reject the request for other reasons, such as the FS being read-only (HDFS), the block size being below the minimum permitted (HDFS), the replication count being out of range (HDFS), quotas on namespace or filesystem being exceeded, reserved names, etc. All rejections SHOULD be IOException or a subclass thereof and MAY be a RuntimeException or subclass. For instance, HDFS may raise an InvalidPathException.


FS' where :
   FS'.Files[p] == []
   ancestors(p) subset-of FS'.Directories

result = FSDataOutputStream

A zero byte file MUST exist at the end of the specified path, visible to all.

The updated (valid) FileSystem MUST contain all the parent directories of the path, as created by mkdirs(parent(p)).

The result is FSDataOutputStream, which through its operations may generate new filesystem states with updated values of FS.Files[p]

The behavior of the returned stream is covered in Output.

Implementation Notes

  • Some implementations split the create into a check for the file existing from the actual creation. This means the operation is NOT atomic: it is possible for clients creating files with overwrite==true to fail if the file is created by another client between the two tests.

  • The S3A and potentially other Object Stores connectors currently don’t change the FS state until the output stream close() operation is completed. This is a significant difference between the behavior of object stores and that of filesystems, as it allows >1 client to create a file with overwrite=false, and potentially confuse file/directory logic. In particular, using create() to acquire an exclusive lock on a file (whoever creates the file without an error is considered the holder of the lock) may not be a safe algorithm to use when working with object stores.

  • Object stores may create an empty file as a marker when a file is created. However, object stores with overwrite=true semantics may not implement this atomically, so creating files with overwrite=false cannot be used as an implicit exclusion mechanism between processes.

  • The Local FileSystem raises a FileNotFoundException when trying to create a file over a directory, hence it is listed as an exception that MAY be raised when this precondition fails.

  • Not covered: symlinks. The resolved path of the symlink is used as the final path argument to the create() operation

FSDataOutputStreamBuilder createFile(Path p)

Make a FSDataOutputStreamBuilder to specify the parameters to create a file.

The behavior of the returned stream is covered in Output.

Implementation Notes

createFile(p) returns a FSDataOutputStreamBuilder only and does not make changes on the filesystem immediately. When build() is invoked on the FSDataOutputStreamBuilder, the builder parameters are verified and create(Path p) is invoked on the underlying filesystem. build() has the same preconditions and postconditions as create(Path p).

  • Similar to create(Path p), files are overwritten by default, unless specified by builder.overwrite(false).
  • Unlike create(Path p), missing parent directories are not created by default, unless specified by builder.recursive().

FSDataOutputStream append(Path p, int bufferSize, Progressable progress)

Implementations without a compliant call SHOULD throw UnsupportedOperationException.


if not exists(FS, p) : raise FileNotFoundException

if not isFile(FS, p) : raise {FileAlreadyExistsException, FileNotFoundException, IOException}


FS' = FS
result = FSDataOutputStream

Return: FSDataOutputStream, which can update the entry FS'.Files[p] by appending data to the existing list.

The behavior of the returned stream is covered in Output.

FSDataOutputStreamBuilder appendFile(Path p)

Make a FSDataOutputStreamBuilder to specify the parameters to append to an existing file.

The behavior of the returned stream is covered in Output.

Implementation Notes

appendFile(p) returns a FSDataOutputStreamBuilder only and does not make change on filesystem immediately. When build() is invoked on the FSDataOutputStreamBuilder, the builder parameters are verified and append() is invoked on the underlying filesystem. build() has the same preconditions and postconditions as append().

FSDataInputStream open(Path f, int bufferSize)

Implementations without a compliant call SHOULD throw UnsupportedOperationException.


if not isFile(FS, p)) : raise {FileNotFoundException, IOException}

This is a critical precondition. Implementations of some FileSystems (e.g. Object stores) could shortcut one round trip by postponing their HTTP GET operation until the first read() on the returned FSDataInputStream. However, much client code does depend on the existence check being performed at the time of the open() operation. Implementations MUST check for the presence of the file at the time of creation. This does not imply that the file and its data is still at the time of the following read() or any successors.


result = FSDataInputStream(0, FS.Files[p])

The result provides access to the byte array defined by FS.Files[p]; whether that access is to the contents at the time the open() operation was invoked, or whether and how it may pick up changes to that data in later states of FS is an implementation detail.

The result MUST be the same for local and remote callers of the operation.

HDFS implementation notes

  1. HDFS MAY throw UnresolvedPathException when attempting to traverse symbolic links

  2. HDFS throws IOException("Cannot open filename " + src) if the path exists in the metadata, but no copies of its blocks can be located; -FileNotFoundException would seem more accurate and useful.

FSDataInputStreamBuilder openFile(Path path)

See openFile().

FSDataInputStreamBuilder openFile(PathHandle)

See openFile().

PathHandle getPathHandle(FileStatus stat, HandleOpt... options)

Implementations without a compliant call MUST throw UnsupportedOperationException


let stat = getFileStatus(Path p)
let FS' where:
  (FS'.Directories, FS.Files', FS'.Symlinks)
  p' in paths(FS') where:
    exists(FS, stat.path) implies exists(FS', p')

The referent of a FileStatus instance, at the time it was resolved, is the same referent as the result of getPathHandle(FileStatus). The PathHandle may be used in subsequent operations to ensure invariants hold between calls.

The options parameter specifies whether a subsequent call e.g., open(PathHandle) will succeed if the referent data or location changed. By default, any modification results in an error. The caller MAY specify relaxations that allow operations to succeed even if the referent exists at a different path and/or its data are changed.

An implementation MUST throw UnsupportedOperationException if it cannot support the semantics specified by the caller. The default set of options are as follows.

Unmoved Moved

Changes to ownership, extended attributes, and other metadata are not required to match the PathHandle. Implementations can extend the set of HandleOpt parameters with custom constraints.


A client specifies that the PathHandle should track the entity across renames using REFERENCE. The implementation MUST throw an UnsupportedOperationException when creating the PathHandle unless failure to resolve the reference implies the entity no longer exists.

A client specifies that the PathHandle should resolve only if the entity is unchanged using PATH. The implementation MUST throw an UnsupportedOperationException when creating the PathHandle unless it can distinguish between an identical entity located subsequently at the same path.


result = PathHandle(p')

Implementation notes

The referent of a PathHandle is the namespace when the FileStatus instance was created, not its state when the PathHandle is created. An implementation MAY reject attempts to create or resolve PathHandle instances that are valid, but expensive to service.

Object stores that implement rename by copying objects MUST NOT claim to support CONTENT and REFERENCE unless the lineage of the object is resolved.

It MUST be possible to serialize a PathHandle instance and reinstantiate it in one or more processes, on another machine, and arbitrarily far into the future without changing its semantics. The implementation MUST refuse to resolve instances if it can no longer guarantee its invariants.

HDFS implementation notes

HDFS does not support PathHandle references to directories or symlinks. Support for CONTENT and REFERENCE looks up files by INode. INodes are not unique across NameNodes, so federated clusters SHOULD include enough metadata in the PathHandle to detect references from other namespaces.

FSDataInputStream open(PathHandle handle, int bufferSize)

Implementations without a compliant call MUST throw UnsupportedOperationException


let fd = getPathHandle(FileStatus stat)
if stat.isdir : raise IOException
let FS' where:
  (FS'.Directories, FS.Files', FS'.Symlinks)
  p' in FS'.Files where:
    FS'.Files[p'] = fd
if not exists(FS', p') : raise InvalidPathHandleException

The implementation MUST resolve the referent of the PathHandle following the constraints specified at its creation by getPathHandle(FileStatus).

Metadata necessary for the FileSystem to satisfy this contract MAY be encoded in the PathHandle.


result = FSDataInputStream(0, FS'.Files[p'])

The stream returned is subject to the constraints of a stream returned by open(Path). Constraints checked on open MAY hold to hold for the stream, but this is not guaranteed.

For example, a PathHandle created with CONTENT constraints MAY return a stream that ignores updates to the file after it is opened, if it was unmodified when open(PathHandle) was resolved.

Implementation notes

An implementation MAY check invariants either at the server or before returning the stream to the client. For example, an implementation may open the file, then verify the invariants in the PathHandle using getFileStatus(Path) to implement CONTENT. This could yield false positives and it requires additional RPC traffic.

boolean delete(Path p, boolean recursive)

Delete a path, be it a file, symbolic link or directory. The recursive flag indicates whether a recursive delete should take place —if unset then a non-empty directory cannot be deleted.

Except in the special case of the root directory, if this API call completed successfully then there is nothing at the end of the path. That is: the outcome is desired. The return flag simply tells the caller whether or not any change was made to the state of the filesystem.

Note: many uses of this method surround it with checks for the return value being false, raising exception if so. For example

if (!fs.delete(path, true)) throw new IOException("Could not delete " + path);

This pattern is not needed. Code SHOULD just call delete(path, recursive) and assume the destination is no longer present —except in the special case of root directories, which will always remain (see below for special coverage of root directories).


A directory with children and recursive == False cannot be deleted

if isDir(FS, p) and not recursive and (children(FS, p) != {}) : raise IOException

(HDFS raises PathIsNotEmptyDirectoryException here.)


Nonexistent path

If the file does not exist the filesystem state does not change

if not exists(FS, p) :
    FS' = FS
    result = False

The result SHOULD be False, indicating that no file was deleted.

Simple File

A path referring to a file is removed, return value: True

if isFile(FS, p) :
    FS' = (FS.Directories, FS.Files - [p], FS.Symlinks)
    result = True
Empty root directory, recursive == False

Deleting an empty root does not change the filesystem state and may return true or false.

if isRoot(p) and children(FS, p) == {} :
    FS ' = FS
    result = (undetermined)

There is no consistent return code from an attempt to delete the root directory.

Implementations SHOULD return true; this avoids code which checks for a false return value from overreacting.

Object Stores: see Object Stores: root directory deletion.

Empty (non-root) directory recursive == False

Deleting an empty directory that is not root will remove the path from the FS and return true.

if isDir(FS, p) and not isRoot(p) and children(FS, p) == {} :
    FS' = (FS.Directories - [p], FS.Files, FS.Symlinks)
    result = True
Recursive delete of non-empty root directory

Deleting a root path with children and recursive==True can generally have three outcomes:

  1. The POSIX model assumes that if the user has the correct permissions to delete everything, they are free to do so (resulting in an empty filesystem).

    if isDir(FS, p) and isRoot(p) and recursive :
        FS' = ({["/"]}, {}, {}, {})
        result = True
  2. HDFS never permits the deletion of the root of a filesystem; the filesystem must be taken offline and reformatted if an empty filesystem is desired.

    if isDir(FS, p) and isRoot(p) and recursive :
        FS' = FS
        result = False
  3. Object Stores: see Object Stores: root directory deletion.

This specification does not recommend any specific action. Do note, however, that the POSIX model assumes that there is a permissions model such that normal users do not have the permission to delete that root directory; it is an action which only system administrators should be able to perform.

Any filesystem client which interacts with a remote filesystem which lacks such a security model, MAY reject calls to delete("/", true) on the basis that it makes it too easy to lose data.

Object Stores: root directory deletion

Some of the object store based filesystem implementations always return false when deleting the root, leaving the state of the store unchanged.

if isRoot(p) :
    FS' = FS
    result = False

This is irrespective of the recursive flag status or the state of the directory.

This is a simplification which avoids the inevitably non-atomic scan and delete of the contents of the store. It also avoids any confusion about whether the operation actually deletes that specific store/container itself, and adverse consequences of the simpler permissions models of stores.

Recursive delete of non-root directory

Deleting a non-root path with children recursive==true removes the path and all descendants

if isDir(FS, p) and not isRoot(p) and recursive :
    FS' where:
        not isDir(FS', p)
        and forall d in descendants(FS, p):
            not isDir(FS', d)
            not isFile(FS', d)
            not isSymlink(FS', d)
    result = True


  • Deleting a file MUST be an atomic action.

  • Deleting an empty directory MUST be an atomic action.

  • A recursive delete of a directory tree MUST be atomic.

Implementation Notes

  • Object Stores and other non-traditional filesystems onto which a directory tree is emulated, tend to implement delete() as recursive listing and entry-by-entry delete operation. This can break the expectations of client applications for O(1) atomic directory deletion, preventing the stores’ use as drop-in replacements for HDFS.

boolean rename(Path src, Path d)

In terms of its specification, rename() is one of the most complex operations within a filesystem.

In terms of its implementation, it is the one with the most ambiguity regarding when to return false versus raising an exception.

Rename includes the calculation of the destination path. If the destination exists and is a directory, the final destination of the rename becomes the destination + the filename of the source path.

let dest = if (isDir(FS, d) and d != src) :
        d + [filename(src)]
    else :


All checks on the destination path MUST take place after the final dest path has been calculated.

Source src must exist:

if not exists(FS, src) : raise FileNotFoundException

dest cannot be a descendant of src:

if isDescendant(FS, src, dest) : raise IOException

This implicitly covers the special case of isRoot(FS, src).

dest must be root, or have a parent that exists:

if not (isRoot(FS, dest) or exists(FS, parent(dest))) : raise IOException

The parent path of a destination must not be a file:

if isFile(FS, parent(dest)) : raise IOException

This implicitly covers all the ancestors of the parent.

There must not be an existing file at the end of the destination path:

if isFile(FS, dest) : raise FileAlreadyExistsException, IOException


Renaming a directory onto itself

Renaming a directory onto itself is no-op; return value is not specified.

In POSIX the result is False; in HDFS the result is True.

if isDir(FS, src) and src == dest :
    FS' = FS
    result = (undefined)
Renaming a file to self

Renaming a file to itself is a no-op; the result is True.

 if isFile(FS, src) and src == dest :
     FS' = FS
     result = True
Renaming a file onto a nonexistent path

Renaming a file where the destination is a directory moves the file as a child of the destination directory, retaining the filename element of the source path.

if isFile(FS, src) and src != dest:
    FS' where:
        not exists(FS', src)
        and exists(FS', dest)
        and data(FS', dest) == data (FS, source)
    result = True
Renaming a directory onto a directory

If src is a directory then all its children will then exist under dest, while the path src and its descendants will no longer exist. The names of the paths under dest will match those under src, as will the contents:

if isDir(FS, src) and isDir(FS, dest) and src != dest :
    FS' where:
        not exists(FS', src)
        and dest in FS'.Directories
        and forall c in descendants(FS, src) :
            not exists(FS', c))
        and forall c in descendants(FS, src) where isDir(FS, c):
            isDir(FS', dest + childElements(src, c)
        and forall c in descendants(FS, src) where not isDir(FS, c):
                data(FS', dest + childElements(s, c)) == data(FS, c)
    result = True
Renaming into a path where the parent path does not exist
  not exists(FS, parent(dest))

There is no consistent behavior here.


The outcome is no change to FileSystem state, with a return value of false.

FS' = FS
result = False

Local Filesystem

The outcome is as a normal rename, with the additional (implicit) feature that the parent directories of the destination also exist.

exists(FS', parent(dest))

S3A FileSystem

The outcome is as a normal rename, with the additional (implicit) feature that the parent directories of the destination then exist: exists(FS', parent(dest))

There is a check for and rejection if the parent(dest) is a file, but no checks for any other ancestors.

Other Filesystems

Other filesystems strictly reject the operation, raising a FileNotFoundException

Concurrency requirements
  • The core operation of rename()—moving one entry in the filesystem to another—MUST be atomic. Some applications rely on this as a way to coordinate access to data.

  • Some FileSystem implementations perform checks on the destination FileSystem before and after the rename. One example of this is ChecksumFileSystem, which provides checksummed access to local data. The entire sequence MAY NOT be atomic.

Implementation Notes

Files open for reading, writing or appending

The behavior of rename() on an open file is unspecified: whether it is allowed, what happens to later attempts to read from or write to the open stream

Renaming a directory onto itself

The return code of renaming a directory onto itself is unspecified.

Destination exists and is a file

Renaming a file atop an existing file is specified as failing, raising an exception.

  • Local FileSystem : the rename succeeds; the destination file is replaced by the source file.

  • HDFS : The rename fails, no exception is raised. Instead the method call simply returns false.

Missing source file

If the source file src does not exist, FileNotFoundException should be raised.

HDFS fails without raising an exception; rename() merely returns false.

FS' = FS
result = false

The behavior of HDFS here should not be considered a feature to replicate. FileContext explicitly changed the behavior to raise an exception, and the retrofitting of that action to the DFSFileSystem implementation is an ongoing matter for debate.

void concat(Path p, Path sources[])

Joins multiple blocks together to create a single file. This is a little-used operation currently implemented only by HDFS.

Implementations without a compliant call SHOULD throw UnsupportedOperationException.


if not exists(FS, p) : raise FileNotFoundException

if sources==[] : raise IllegalArgumentException

All sources MUST be in the same directory:

for s in sources:
    if parent(s) != parent(p) : raise IllegalArgumentException

All block sizes must match that of the target:

for s in sources:
    getBlockSize(FS, s) == getBlockSize(FS, p)

No duplicate paths:

let input = sources + [p]
not (exists i, j: i != j and input[i] == input[j])

HDFS: All source files except the final one MUST be a complete block:

for s in (sources[0:length(sources)-1] + [p]):
    (length(FS, s) mod getBlockSize(FS, p)) == 0


FS' where:
    (data(FS', p) = data(FS, p) + data(FS, sources[0]) + ... + data(FS, sources[length(sources)-1]))
    for s in sources: not exists(FS', s)

HDFS’s restrictions may be an implementation detail of how it implements concat by changing the inode references to join them together in a sequence. As no other filesystem in the Hadoop core codebase implements this method, there is no way to distinguish implementation detail from specification.

boolean truncate(Path p, long newLength)

Truncate file p to the specified newLength.

Implementations without a compliant call SHOULD throw UnsupportedOperationException.


if not exists(FS, p) : raise FileNotFoundException

if isDir(FS, p) : raise {FileNotFoundException, IOException}

if newLength < 0 || newLength > len(FS.Files[p]) : raise HadoopIllegalArgumentException

HDFS: The source file MUST be closed. Truncate cannot be performed on a file, which is open for writing or appending.


len(FS'.Files[p]) = newLength

Return: true, if truncation is finished and the file can be immediately opened for appending, or false otherwise.

HDFS: HDFS returns false to indicate that a background process of adjusting the length of the last block has been started, and clients should wait for it to complete before they can proceed with further file updates.


If an input stream is open when truncate() occurs, the outcome of read operations related to the part of the file being truncated is undefined.

boolean copyFromLocalFile(boolean delSrc, boolean overwrite, Path src, Path dst)

The source file or directory at src is on the local disk and is copied into the file system at destination dst. If the source must be deleted after the move then delSrc flag must be set to TRUE. If destination already exists, and the destination contents must be overwritten then overwrite flag must be set to TRUE.


Source and destination must be different

if src = dest : raise FileExistsException

Destination and source must not be descendants of one another

if isDescendant(src, dest) or isDescendant(dest, src) : raise IOException

The source file or directory must exist locally:

if not exists(LocalFS, src) : raise FileNotFoundException

Directories cannot be copied into files regardless to what the overwrite flag is set to:

if isDir(LocalFS, src) and isFile(FS, dst) : raise PathExistsException

For all cases, except the one for which the above precondition throws, the overwrite flag must be set to TRUE for the operation to succeed if destination exists. This will also overwrite any files / directories at the destination:

if exists(FS, dst) and not overwrite : raise PathExistsException

Determining the final name of the copy

Given a base path on the source base and a child path child where base is in ancestors(child) + child:

def final_name(base, child, dest):
    if base == child:
        return dest
        return dest + childElements(base, child)

Outcome where source is a file isFile(LocalFS, src)

For a file, data at destination becomes that of the source. All ancestors are directories.

if isFile(LocalFS, src) and (not exists(FS, dest) or (exists(FS, dest) and overwrite)):
    FS' = FS where:
        FS'.Files[dest] = LocalFS.Files[src]
        FS'.Directories = FS.Directories + ancestors(FS, dest)
    LocalFS' = LocalFS where
        not delSrc or (delSrc = true and delete(LocalFS, src, false))
else if isFile(LocalFS, src) and isDir(FS, dest):
    FS' = FS where:
        let d = final_name(src, dest)
        FS'.Files[d] = LocalFS.Files[src]
    LocalFS' = LocalFS where:
        not delSrc or (delSrc = true and delete(LocalFS, src, false))

There are no expectations that the file changes are atomic for both local LocalFS and remote FS.

Outcome where source is a directory isDir(LocalFS, src)

if isDir(LocalFS, src) and (isFile(FS, dest) or isFile(FS, dest + childElements(src))):
    raise FileAlreadyExistsException
else if isDir(LocalFS, src):
    if exists(FS, dest):
        dest' = dest + childElements(src)
        if exists(FS, dest') and not overwrite:
            raise PathExistsException
        dest' = dest

    FS' = FS where:
        forall c in descendants(LocalFS, src):
            not exists(FS', final_name(c)) or overwrite
        and forall c in descendants(LocalFS, src) where isDir(LocalFS, c):
            FS'.Directories = FS'.Directories + (dest' + childElements(src, c))
        and forall c in descendants(LocalFS, src) where isFile(LocalFS, c):
            FS'.Files[final_name(c, dest')] = LocalFS.Files[c]
    LocalFS' = LocalFS where
        not delSrc or (delSrc = true and delete(LocalFS, src, true))

There are no expectations of operation isolation / atomicity. This means files can change in source or destination while the operation is executing. No guarantees are made for the final state of the file or directory after a copy other than it is best effort. E.g.: when copying a directory, one file can be moved from source to destination but there’s nothing stopping the new file at destination being updated while the copy operation is still in place.


The default HDFS implementation, is to recurse through each file and folder, found at src, and copy them sequentially to their final destination (relative to dst).

Object store based file systems should be mindful of what limitations arise from the above implementation and could take advantage of parallel uploads and possible re-ordering of files copied into the store to maximize throughput.

interface RemoteIterator

The RemoteIterator interface is used as a remote-access equivalent to java.util.Iterator, allowing the caller to iterate through a finite sequence of remote data elements.

The core differences are

  1. Iterator’s optional void remove() method is not supported.
  2. For those methods which are supported, IOException exceptions may be raised.
public interface RemoteIterator<E> {
  boolean hasNext() throws IOException;
  E next() throws IOException;

The basic view of the interface is that hasNext() being true implies that next() will successfully return the next entry in the list:

while hasNext(): next()

Equally, a successful call to next() implies that had hasNext() been invoked prior to the call to next(), it would have been true.

boolean elementAvailable = hasNext();
try {
  assert elementAvailable;
} catch (NoSuchElementException e) {
  assert !elementAvailable

The next() operator MUST iterate through the list of available results, even if no calls to hasNext() are made.

That is, it is possible to enumerate the results through a loop which only terminates when a NoSuchElementException exception is raised.

try {
  while (true) {
} catch (NoSuchElementException ignored) {
  // the end of the list has been reached

The output of the iteration is equivalent to the loop

while (iterator.hasNext()) {

As raising exceptions is an expensive operation in JVMs, the while(hasNext()) loop option is more efficient. (see also Concurrency and the Remote Iterator for a discussion on this topic).

Implementors of the interface MUST support both forms of iterations; authors of tests SHOULD verify that both iteration mechanisms work.

The iteration is required to return a finite sequence; both forms of loop MUST ultimately terminate. All implementations of the interface in the Hadoop codebase meet this requirement; all consumers assume that it holds.

boolean hasNext()

Returns true if-and-only-if a subsequent single call to next() would return an element rather than raise an exception.



result = True ==> next() will succeed.
result = False ==> next() will raise an exception

Multiple calls to hasNext(), without any intervening next() calls, MUST return the same value.

boolean has1 = iterator.hasNext();
boolean has2 = iterator.hasNext();
assert has1 == has2;

E next()

Return the next element in the iteration.


hasNext() else raise java.util.NoSuchElementException


result = the next element in the iteration

Repeated calls to next() return subsequent elements in the sequence, until the entire sequence has been returned.

Concurrency and the Remote Iterator

The primary use of RemoteIterator in the filesystem APIs is to list files on (possibly remote) filesystems. These filesystems are invariably accessed concurrently; the state of the filesystem MAY change between a hasNext() probe and the invocation of the next() call.

During iteration through a RemoteIterator, if the directory is deleted on remote filesystem, then hasNext() or next() call may throw FileNotFoundException.

Accordingly, a robust iteration through a RemoteIterator would catch and discard NoSuchElementException exceptions raised during the process, which could be done through the while(true) iteration example above, or through a hasNext()/next() sequence with an outer try/catch clause to catch a NoSuchElementException alongside other exceptions which may be raised during a failure (for example, a FileNotFoundException)

try {
  while (iterator.hasNext()) {
} catch (NoSuchElementException ignored) {
  // the end of the list has been reached

It is notable that this is not done in the Hadoop codebase. This does not imply that robust loops are not recommended —more that the concurrency problems were not considered during the implementation of these loops.

interface StreamCapabilities

The StreamCapabilities provides a way to programmatically query the capabilities that OutputStream, InputStream, or other FileSystem class supports.

public interface StreamCapabilities {
  boolean hasCapability(String capability);

boolean hasCapability(capability)

Return true iff the OutputStream, InputStream, or other FileSystem class has the desired capability.

The caller can query the capabilities of a stream using a string value. Here is a table of possible string values:

String Constant Implements Description
hflush HFLUSH Syncable Flush out the data in client’s user buffer. After the return of this call, new readers will see the data.
hsync HSYNC Syncable Flush out the data in client’s user buffer all the way to the disk device (but the disk may have it in its cache). Similar to POSIX fsync.
in:readahead READAHEAD CanSetReadahead Set the readahead on the input stream.
dropbehind DROPBEHIND CanSetDropBehind Drop the cache.
in:unbuffer UNBUFFER CanUnbuffer Reduce the buffering on the input stream.

Etag probes through the interface EtagSource

FileSystem implementations MAY support querying HTTP etags from FileStatus entries. If so, the requirements are as follows

Etag support MUST BE across all list/getFileStatus() calls.

That is: when adding etag support, all operations which return FileStatus or ListLocatedStatus entries MUST return subclasses which are instances of EtagSource.

FileStatus instances MUST have etags whenever the remote store provides them.

To support etags, they MUST BE to be provided in both getFileStatus() and list calls.

Implementors note: the core APIs which MUST BE overridden to achieve this are as follows:

FileStatus getFileStatus(Path)
FileStatus[] listStatus(Path)
RemoteIterator<FileStatus> listStatusIterator(Path)
RemoteIterator<LocatedFileStatus> listFiles([Path, boolean)

Etags of files MUST BE Consistent across all list/getFileStatus operations.

The value of EtagSource.getEtag() MUST be the same for list* queries which return etags for calls of getFileStatus() for the specific object.

((EtagSource)getFileStatus(path)).getEtag() == ((EtagSource)listStatus(path)[0]).getEtag()

Similarly, the same value MUST BE returned for listFiles(), listStatusIncremental() of the path and when listing the parent path, of all files in the listing.

Etags MUST BE different for different file contents.

Two different arrays of data written to the same path MUST have different etag values when probed. This is a requirement of the HTTP specification.

Etags of files SHOULD BE preserved across rename operations

After a file is renamed, the value of ((EtagSource)getFileStatus(dest)).getEtag() SHOULD be the same as the value of ((EtagSource)getFileStatus(source)).getEtag() was before the rename took place.

This is an implementation detail of the store; it does not hold for AWS S3.

If and only if the store consistently meets this requirement, the filesystem SHOULD declare in hasPathCapability() that it supports

Directories MAY have etags

Directory entries MAY return etags in listing/probe operations; these entries MAY be preserved across renames.

Equally, directory entries MAY NOT provide such entries, MAY NOT preserve them acrosss renames, and MAY NOT guarantee consistency over time.

Note: special mention of the root path “/”. As that isn’t a real “directory”, nobody should expect it to have an etag.

All etag-aware FileStatus subclass MUST BE Serializable; MAY BE Writable

The base FileStatus class implements Serializable and Writable and marshalls its fields appropriately.

Subclasses MUST support java serialization (Some Apache Spark applications use it), preserving the etag. This is a matter of making the etag field non-static and adding a serialVersionUID.

The Writable support was used for marshalling status data over Hadoop IPC calls; in Hadoop 3 that is implemented through org/apache/hadoop/fs/protocolPB/PBHelper.javaand the methods deprecated. Subclasses MAY override the deprecated methods to add etag marshalling. However -but there is no expectation of this and such marshalling is unlikely to ever take place.

Appropriate etag Path Capabilities SHOULD BE declared

  1. hasPathCapability(path, "fs.capability.etags.available") MUST return true iff the filesystem returns valid (non-empty etags) on file status/listing operations.
  2. hasPathCapability(path, "fs.capability.etags.consistent.across.rename") MUST return true if and only if etags are preserved across renames.

Non-requirements of etag support

  • There is no requirement/expectation that FileSystem.getFileChecksum(Path) returns a checksum value related to the etag of an object, if any value is returned.
  • If the same data is uploaded to the twice to the same or a different path, the etag of the second upload MAY NOT match that of the first upload.