Following system colour scheme Selected dark colour scheme Selected light colour scheme

Python Enhancement Proposals

PEP 3116 – New I/O

Daniel Stutzbach <daniel at>, Guido van Rossum <guido at>, Mike Verdone <mike.verdone at>
Standards Track

Table of Contents

Rationale and Goals

Python allows for a variety of stream-like (a.k.a. file-like) objects that can be used via read() and write() calls. Anything that provides read() and write() is stream-like. However, more exotic and extremely useful functions like readline() or seek() may or may not be available on every stream-like object. Python needs a specification for basic byte-based I/O streams to which we can add buffering and text-handling features.

Once we have a defined raw byte-based I/O interface, we can add buffering and text handling layers on top of any byte-based I/O class. The same buffering and text handling logic can be used for files, sockets, byte arrays, or custom I/O classes developed by Python programmers. Developing a standard definition of a stream lets us separate stream-based operations like read() and write() from implementation specific operations like fileno() and isatty(). It encourages programmers to write code that uses streams as streams and not require that all streams support file-specific or socket-specific operations.

The new I/O spec is intended to be similar to the Java I/O libraries, but generally less confusing. Programmers who don’t want to muck about in the new I/O world can expect that the open() factory method will produce an object backwards-compatible with old-style file objects.


The Python I/O Library will consist of three layers: a raw I/O layer, a buffered I/O layer, and a text I/O layer. Each layer is defined by an abstract base class, which may have multiple implementations. The raw I/O and buffered I/O layers deal with units of bytes, while the text I/O layer deals with units of characters.

Raw I/O

The abstract base class for raw I/O is RawIOBase. It has several methods which are wrappers around the appropriate operating system calls. If one of these functions would not make sense on the object, the implementation must raise an IOError exception. For example, if a file is opened read-only, the .write() method will raise an IOError. As another example, if the object represents a socket, then .seek(), .tell(), and .truncate() will raise an IOError. Generally, a call to one of these functions maps to exactly one operating system call.

.read(n: int) -> bytes
Read up to n bytes from the object and return them. Fewer than n bytes may be returned if the operating system call returns fewer than n bytes. If 0 bytes are returned, this indicates end of file. If the object is in non-blocking mode and no bytes are available, the call returns None.

.readinto(b: bytes) -> int

Read up to len(b) bytes from the object and stores them in b, returning the number of bytes read. Like .read, fewer than len(b) bytes may be read, and 0 indicates end of file. None is returned if a non-blocking object has no bytes available. The length of b is never changed.

.write(b: bytes) -> int

Returns number of bytes written, which may be < len(b).

.seek(pos: int, whence: int = 0) -> int

.tell() -> int

.truncate(n: int = None) -> int

.close() -> None

Additionally, it defines a few other methods:

.readable() -> bool
Returns True if the object was opened for reading, False otherwise. If False, .read() will raise an IOError if called.

.writable() -> bool

Returns True if the object was opened for writing, False otherwise. If False, .write() and .truncate() will raise an IOError if called.

.seekable() -> bool

Returns True if the object supports random access (such as disk files), or False if the object only supports sequential access (such as sockets, pipes, and ttys). If False, .seek(), .tell(), and .truncate() will raise an IOError if called.

.__enter__() -> ContextManager

Context management protocol. Returns self.

.__exit__(...) -> None

Context management protocol. Same as .close().

If and only if a RawIOBase implementation operates on an underlying file descriptor, it must additionally provide a .fileno() member function. This could be defined specifically by the implementation, or a mix-in class could be used (need to decide about this).

.fileno() -> int
Returns the underlying file descriptor (an integer)

Initially, three implementations will be provided that implement the RawIOBase interface: FileIO, SocketIO (in the socket module), and ByteIO. Each implementation must determine whether the object supports random access as the information provided by the user may not be sufficient (consider open("/dev/tty", "rw") or open("/tmp/named-pipe", "rw")). As an example, FileIO can determine this by calling the seek() system call; if it returns an error, the object does not support random access. Each implementation may provided additional methods appropriate to its type. The ByteIO object is analogous to Python 2’s cStringIO library, but operating on the new bytes type instead of strings.

Buffered I/O

The next layer is the Buffered I/O layer which provides more efficient access to file-like objects. The abstract base class for all Buffered I/O implementations is BufferedIOBase, which provides similar methods to RawIOBase:

.read(n: int = -1) -> bytes
Returns the next n bytes from the object. It may return fewer than n bytes if end-of-file is reached or the object is non-blocking. 0 bytes indicates end-of-file. This method may make multiple calls to to gather the bytes, or may make no calls to if all of the needed bytes are already buffered.

.readinto(b: bytes) -> int

.write(b: bytes) -> int

Write b bytes to the buffer. The bytes are not guaranteed to be written to the Raw I/O object immediately; they may be buffered. Returns len(b).

.seek(pos: int, whence: int = 0) -> int

.tell() -> int

.truncate(pos: int = None) -> int

.flush() -> None

.close() -> None

.readable() -> bool

.writable() -> bool

.seekable() -> bool

.__enter__() -> ContextManager

.__exit__(...) -> None

Additionally, the abstract base class provides one member variable:

A reference to the underlying RawIOBase object.

The BufferedIOBase methods signatures are mostly identical to that of RawIOBase (exceptions: write() returns None, read()’s argument is optional), but may have different semantics. In particular, BufferedIOBase implementations may read more data than requested or delay writing data using buffers. For the most part, this will be transparent to the user (unless, for example, they open the same file through a different descriptor). Also, raw reads may return a short read without any particular reason; buffered reads will only return a short read if EOF is reached; and raw writes may return a short count (even when non-blocking I/O is not enabled!), while buffered writes will raise IOError when not all bytes could be written or buffered.

There are four implementations of the BufferedIOBase abstract base class, described below.


The BufferedReader implementation is for sequential-access read-only objects. Its .flush() method is a no-op.


The BufferedWriter implementation is for sequential-access write-only objects. Its .flush() method forces all cached data to be written to the underlying RawIOBase object.


The BufferedRWPair implementation is for sequential-access read-write objects such as sockets and ttys. As the read and write streams of these objects are completely independent, it could be implemented by simply incorporating a BufferedReader and BufferedWriter instance. It provides a .flush() method that has the same semantics as a BufferedWriter’s .flush() method.


The BufferedRandom implementation is for all random-access objects, whether they are read-only, write-only, or read-write. Compared to the previous classes that operate on sequential-access objects, the BufferedRandom class must contend with the user calling .seek() to reposition the stream. Therefore, an instance of BufferedRandom must keep track of both the logical and true position within the object. It provides a .flush() method that forces all cached write data to be written to the underlying RawIOBase object and all cached read data to be forgotten (so that future reads are forced to go back to the disk).

Q: Do we want to mandate in the specification that switching between reading and writing on a read-write object implies a .flush()? Or is that an implementation convenience that users should not rely on?

For a read-only BufferedRandom object, .writable() returns False and the .write() and .truncate() methods throw IOError.

For a write-only BufferedRandom object, .readable() returns False and the .read() method throws IOError.

Text I/O

The text I/O layer provides functions to read and write strings from streams. Some new features include universal newlines and character set encoding and decoding. The Text I/O layer is defined by a TextIOBase abstract base class. It provides several methods that are similar to the BufferedIOBase methods, but operate on a per-character basis instead of a per-byte basis. These methods are:

.read(n: int = -1) -> str

.write(s: str) -> int

.tell() -> object

Return a cookie describing the current file position. The only supported use for the cookie is with .seek() with whence set to 0 (i.e. absolute seek).

.seek(pos: object, whence: int = 0) -> int

Seek to position pos. If pos is non-zero, it must be a cookie returned from .tell() and whence must be zero.

.truncate(pos: object = None) -> int

Like BufferedIOBase.truncate(), except that pos (if not None) must be a cookie previously returned by .tell().

Unlike with raw I/O, the units for .seek() are not specified - some implementations (e.g. StringIO) use characters and others (e.g. TextIOWrapper) use bytes. The special case for zero is to allow going to the start or end of a stream without a prior .tell(). An implementation could include stream encoder state in the cookie returned from .tell().

TextIOBase implementations also provide several methods that are pass-throughs to the underlying BufferedIOBase objects:

.flush() -> None

.close() -> None

.readable() -> bool

.writable() -> bool

.seekable() -> bool

TextIOBase class implementations additionally provide the following methods:

.readline() -> str
Read until newline or EOF and return the line, or "" if EOF hit immediately.

.__iter__() -> Iterator

Returns an iterator that returns lines from the file (which happens to be self).

.next() -> str

Same as readline() except raises StopIteration if EOF hit immediately.

Two implementations will be provided by the Python library. The primary implementation, TextIOWrapper, wraps a Buffered I/O object. Each TextIOWrapper object has a property named “.buffer” that provides a reference to the underlying BufferedIOBase object. Its initializer has the following signature:

.__init__(self, buffer, encoding=None, errors=None, newline=None, line_buffering=False)
buffer is a reference to the BufferedIOBase object to be wrapped with the TextIOWrapper.

encoding refers to an encoding to be used for translating between the byte-representation and character-representation. If it is None, then the system’s locale setting will be used as the default.

errors is an optional string indicating error handling. It may be set whenever encoding may be set. It defaults to 'strict'.

newline can be None, '', '\n', '\r', or '\r\n'; all other values are illegal. It controls the handling of line endings. It works as follows:

  • On input, if newline is None, universal newlines mode is enabled. Lines in the input can end in '\n', '\r', or '\r\n', and these are translated into '\n' before being returned to the caller. If it is '', universal newline mode is enabled, but line endings are returned to the caller untranslated. If it has any of the other legal values, input lines are only terminated by the given string, and the line ending is returned to the caller untranslated. (In other words, translation to '\n' only occurs if newline is None.)
  • On output, if newline is None, any '\n' characters written are translated to the system default line separator, os.linesep. If newline is '', no translation takes place. If newline is any of the other legal values, any '\n' characters written are translated to the given string. (Note that the rules guiding translation are different for output than for input.)

line_buffering, if True, causes write() calls to imply a flush() if the string written contains at least one '\n' or '\r' character. This is set by open() when it detects that the underlying stream is a TTY device, or when a buffering argument of 1 is passed.

Further notes on the newline parameter:

  • '\r' support is still needed for some OSX applications that produce files using '\r' line endings; Excel (when exporting to text) and Adobe Illustrator EPS files are the most common examples.
  • If translation is enabled, it happens regardless of which method is called for reading or writing. For example, will always produce the same result as ''.join(f.readlines()).
  • If universal newlines without translation are requested on input (i.e. newline=''), if a system read operation returns a buffer ending in '\r', another system read operation is done to determine whether it is followed by '\n' or not. In universal newlines mode with translation, the second system read operation may be postponed until the next read request, and if the following system read operation returns a buffer starting with '\n', that character is simply discarded.

Another implementation, StringIO, creates a file-like TextIO implementation without an underlying Buffered I/O object. While similar functionality could be provided by wrapping a BytesIO object in a TextIOWrapper, the StringIO object allows for much greater efficiency as it does not need to actually performing encoding and decoding. A String I/O object can just store the encoded string as-is. The StringIO object’s __init__ signature takes an optional string specifying the initial value; the initial position is always 0. It does not support encodings or newline translations; you always read back exactly the characters you wrote.

Unicode encoding/decoding Issues

We should allow changing the encoding and error-handling setting later. The behavior of Text I/O operations in the face of Unicode problems and ambiguities (e.g. diacritics, surrogates, invalid bytes in an encoding) should be the same as that of the unicode encode()/decode() methods. UnicodeError may be raised.

Implementation note: we should be able to reuse much of the infrastructure provided by the codecs module. If it doesn’t provide the exact APIs we need, we should refactor it to avoid reinventing the wheel.

Non-blocking I/O

Non-blocking I/O is fully supported on the Raw I/O level only. If a raw object is in non-blocking mode and an operation would block, then .read() and .readinto() return None, while .write() returns 0. In order to put an object in non-blocking mode, the user must extract the fileno and do it by hand.

At the Buffered I/O and Text I/O layers, if a read or write fails due a non-blocking condition, they raise an IOError with errno set to EAGAIN.

Originally, we considered propagating up the Raw I/O behavior, but many corner cases and problems were raised. To address these issues, significant changes would need to have been made to the Buffered I/O and Text I/O layers. For example, what should .flush() do on a Buffered non-blocking object? How would the user instruct the object to “Write as much as you can from your buffer, but don’t block”? A non-blocking .flush() that doesn’t necessarily flush all available data is counter-intuitive. Since non-blocking and blocking objects would have such different semantics at these layers, it was agreed to abandon efforts to combine them into a single type.

The open() Built-in Function

The open() built-in function is specified by the following pseudo-code:

def open(filename, mode="r", buffering=None, *,
         encoding=None, errors=None, newline=None):
    assert isinstance(filename, (str, int))
    assert isinstance(mode, str)
    assert buffering is None or isinstance(buffering, int)
    assert encoding is None or isinstance(encoding, str)
    assert newline in (None, "", "\n", "\r", "\r\n")
    modes = set(mode)
    if modes - set("arwb+t") or len(mode) > len(modes):
        raise ValueError("invalid mode: %r" % mode)
    reading = "r" in modes
    writing = "w" in modes
    binary = "b" in modes
    appending = "a" in modes
    updating = "+" in modes
    text = "t" in modes or not binary
    if text and binary:
        raise ValueError("can't have text and binary mode at once")
    if reading + writing + appending > 1:
        raise ValueError("can't have read/write/append mode at once")
    if not (reading or writing or appending):
        raise ValueError("must have exactly one of read/write/append mode")
    if binary and encoding is not None:
        raise ValueError("binary modes doesn't take an encoding arg")
    if binary and errors is not None:
        raise ValueError("binary modes doesn't take an errors arg")
    if binary and newline is not None:
        raise ValueError("binary modes doesn't take a newline arg")
    # XXX Need to spec the signature for FileIO()
    raw = FileIO(filename, mode)
    line_buffering = (buffering == 1 or buffering is None and raw.isatty())
    if line_buffering or buffering is None:
        buffering = 8*1024  # International standard buffer size
        # XXX Try setting it to fstat().st_blksize
    if buffering < 0:
        raise ValueError("invalid buffering size")
    if buffering == 0:
        if binary:
            return raw
        raise ValueError("can't have unbuffered text I/O")
    if updating:
        buffer = BufferedRandom(raw, buffering)
    elif writing or appending:
        buffer = BufferedWriter(raw, buffering)
        assert reading
        buffer = BufferedReader(raw, buffering)
    if binary:
        return buffer
    assert text
    return TextIOWrapper(buffer, encoding, errors, newline, line_buffering)


Last modified: 2023-09-09 17:39:29 GMT