尚未翻译

本页面尚未翻译成中文，因此显示英文原文。帮助翻译

selectors Module Complexity¶

The selectors module wraps the select module's primitives - select(), poll(), epoll, /dev/poll and kqueue - behind one interface: register file objects with the events to watch, then ask which of them are ready. Registration is a dict entry keyed by file descriptor, plus at most a couple of system calls; the cost that differs between the implementations is select() itself.

n is the file objects registered with a selector, and k is the (key, events) pairs one select() call returns - one per ready file object, or two for a file kqueue reports ready for reading and writing separately. The bounds price the work of one call, not the time it spends waiting for a descriptor to become ready, and treat a system call that adds, changes or removes one descriptor as O(1). A file object is an integer descriptor or anything with a fileno() method, which every method that takes one calls to find its registration.

Complexity Reference¶

Selector classes¶

Operation	Time	Space	Notes
`selectors.DefaultSelector()`	O(1)	O(1)	An alias chosen at import: the first of kqueue, epoll, `/dev/poll`, poll and select that works; `EpollSelector` on Linux. Where it is `DevpollSelector`, it costs what that costs
`selectors.SelectSelector()`	O(1)	O(1)	Available everywhere; on Unix, descriptors must be below `FD_SETSIZE` (usually 1024) or `select()` raises `ValueError`
`selectors.PollSelector()`	O(1)	O(1)	Unix
`selectors.EpollSelector()`	O(1)	O(1)	Linux; opens an epoll descriptor
`selectors.KqueueSelector()`	O(1)	O(1)	BSD and macOS; opens a kqueue descriptor
`selectors.DevpollSelector()`	O(1)	O(L)	Solaris; L = the process's open-file limit, one slot per descriptor reserved up front
`EpollSelector.fileno()`, `KqueueSelector.fileno()`, `DevpollSelector.fileno()`	O(1)	O(1)	The kernel object's own descriptor

BaseSelector¶

Operation	Time	Space	Notes
`selectors.BaseSelector`	O(1)	O(1)	Abstract; a subclass supplies `register`, `unregister`, `select` and `get_map`, and inherits `modify` as `unregister` plus `register`
`BaseSelector.register(fileobj, events, data=None)`	O(1)	O(1)	Returns the new `SelectorKey`; `KeyError` if the descriptor is already registered, `ValueError` if `events` is empty or has unknown bits
`BaseSelector.unregister(fileobj)`	O(1)	O(1)	O(n) once `fileobj` is closed: when `fileno()` fails, every registration is scanned for the object
`BaseSelector.modify(fileobj, events, data=None)`	O(1)	O(1)	Returns the updated `SelectorKey`; changing only `data` makes no system call
`BaseSelector.select(timeout=None)`	O(n) or O(k)	O(k) to O(n)	Returns k `(key, events)` pairs; the cost is the implementation's, below. `timeout <= 0` polls and `None` blocks
`BaseSelector.get_key(fileobj)`	O(1)	O(1)	`KeyError` if not registered, `RuntimeError` once the selector is closed
`BaseSelector.get_map()`	O(1)	O(1)	The same read-only view on every call, not a copy; `None` once closed
`len(mapping)`, `mapping[fileobj]`, `fileobj in mapping`	O(1)	O(1)	On the view `get_map()` returns
Iterating the view	O(n)	O(1)	Yields file descriptors, not file objects; `.values()` yields the keys
`BaseSelector.close()`, leaving a `with` block	O(n)	O(1)	Drops every registration and closes the kernel object, if there is one; the registered files stay open

select() by implementation¶

Operation	Time	Space	Notes
`SelectSelector.select(timeout=None)`	O(n)	O(n)	Every registered descriptor is copied into the `select()` call and checked on return
`PollSelector.select(timeout=None)`	O(n)	O(k)	The kernel walks the whole n-entry array the selector keeps between calls; the first call after registrations change rebuilds it, resizing it to n
`EpollSelector.select(timeout=None)`	O(k)	O(n)	The kernel hands back only ready descriptors, but each call allocates a result buffer sized for all n
`KqueueSelector.select(timeout=None)`	O(k)	O(n)	As epoll
`DevpollSelector.select(timeout=None)`	O(k)	O(k)	Reuses the array allocated at construction

SelectorKey and constants¶

Operation	Time	Space	Notes
`selectors.SelectorKey`	O(1)	O(1)	A named tuple: `modify` replaces a registration's key rather than changing it in place
`SelectorKey.fileobj`, `SelectorKey.fd`, `SelectorKey.events`, `SelectorKey.data`	O(1)	O(1)	What was registered, its descriptor, the event mask, and the attached object
`selectors.EVENT_READ`, `selectors.EVENT_WRITE`	O(1)	O(1)	Bit flags, combined with bitwise or to watch both

Choosing a Selector¶

DefaultSelector is not a class of its own but whichever implementation the platform supports best, and all of them share the BaseSelector interface. Use it as a context manager so the kernel object is released.

import selectors

assert selectors.DefaultSelector in (
    selectors.SelectSelector,
    getattr(selectors, 'PollSelector', None),
    getattr(selectors, 'EpollSelector', None),
    getattr(selectors, 'DevpollSelector', None),
    getattr(selectors, 'KqueueSelector', None),
)

with selectors.DefaultSelector() as sel:  # O(1)
    assert isinstance(sel, selectors.BaseSelector)
    assert sel.get_map() is not None
assert sel.get_map() is None  # O(n) close on exit

Registering File Objects¶

register() stores a SelectorKey under the file's descriptor and returns it. A descriptor can be registered once; to watch different events on it, modify() the registration.

import selectors
import socket

left, right = socket.socketpair()

with selectors.DefaultSelector() as sel:
    key = sel.register(left, selectors.EVENT_READ, data='left')  # O(1)
    assert key.fileobj is left and key.fd == left.fileno()
    assert sel.get_key(left) is key  # O(1)

    try:
        sel.register(left, selectors.EVENT_WRITE)
    except KeyError as error:
        assert 'already registered' in str(error)
    else:
        raise AssertionError('a descriptor was registered twice')

    try:
        sel.register(right, 0)
    except ValueError as error:
        assert 'Invalid events' in str(error)
    else:
        raise AssertionError('an empty event mask was accepted')

    both = selectors.EVENT_READ | selectors.EVENT_WRITE
    changed = sel.modify(left, both, data='left')  # O(1)
    assert changed.events == both and key.events == selectors.EVENT_READ
    renamed = sel.modify(left, both, data='renamed')  # O(1) - no system call
    assert sel.get_key(left) is renamed

    assert sel.unregister(left) is renamed  # O(1)

left.close()
right.close()

What select() Returns¶

select() returns a list of (key, events) pairs for the ready file objects, where events is the ready events masked by what was registered. Iterating the result costs O(k) whatever the implementation. kqueue reports reading and writing as separate pairs, so a file watched for both can appear twice.

import selectors
import socket

left, right = socket.socketpair()

with selectors.DefaultSelector() as sel:
    sel.register(left, selectors.EVENT_READ, data='reader')

    assert sel.select(timeout=0) == []  # nothing to read yet

    right.send(b'ping')
    ready = sel.select(timeout=0)  # O(k) for epoll and kqueue, O(n) for poll and select
    assert len(ready) == 1
    key, events = ready[0]
    assert key.data == 'reader' and events == selectors.EVENT_READ
    assert key.fileobj.recv(4) == b'ping'

left.close()
right.close()

How select() Scales¶

select() and poll() hand the kernel every registered descriptor on every call and check each on return, so a call costs O(n) however few are ready. epoll, kqueue and /dev/poll keep the interest set in the kernel and return only the ready descriptors, so a call costs O(k). For many mostly idle connections that is the difference that matters, and it is why DefaultSelector prefers them.

import os
import selectors

pipes = [os.pipe() for _ in range(200)]
os.write(pipes[-1][1], b'x')  # one of 200 is ready

candidates = [selectors.SelectSelector]
for name in ('PollSelector', 'EpollSelector', 'KqueueSelector', 'DevpollSelector'):
    if hasattr(selectors, name):
        candidates.append(getattr(selectors, name))

for cls in candidates:
    with cls() as sel:
        for r, _ in pipes:
            sel.register(r, selectors.EVENT_READ)  # O(1) each
        ready = sel.select(timeout=0)  # O(n) for select/poll, O(k) for the rest
        assert [key.fd for key, _ in ready] == [pipes[-1][0]]

for r, w in pipes:
    os.close(r)
    os.close(w)

The kernel-side selectors are not free of n: epoll and kqueue allocate a result buffer sized for every registration on each call, so their select() is O(k) in time and O(n) in memory.

Unregistering Before Closing¶

Registrations are found by descriptor. Once a file is closed, its fileno() fails, and unregister(), modify() and get_key() fall back to scanning every registration for the object itself - O(n) per call. Unregister first.

import selectors
import socket

with selectors.DefaultSelector() as sel:
    left, right = socket.socketpair()
    sel.register(left, selectors.EVENT_READ)

    sel.unregister(left)  # O(1) - fileno() still works
    left.close()

    sel.register(right, selectors.EVENT_READ)
    right.close()
    assert right.fileno() == -1
    sel.unregister(right)  # O(n) - found by scanning every registration
    assert len(sel.get_map()) == 0

The Registration Map¶

get_map() returns a live, read-only view of the registrations, not a snapshot, so holding it costs nothing and it always reflects the current state. Lookups take a file object; iteration yields descriptors.

import selectors
import socket

left, right = socket.socketpair()

with selectors.DefaultSelector() as sel:
    mapping = sel.get_map()  # O(1)
    assert sel.get_map() is mapping

    sel.register(left, selectors.EVENT_READ, data='left')
    sel.register(right, selectors.EVENT_WRITE, data='right')
    assert len(mapping) == 2  # O(1) - the view sees the new entries
    assert mapping[left].data == 'left' and right in mapping  # O(1)
    assert set(mapping) == {left.fileno(), right.fileno()}  # O(n)
    assert {key.data for key in mapping.values()} == {'left', 'right'}  # O(n)

sel.close()
assert sel.get_map() is None
try:
    sel.get_key(left)
except RuntimeError as error:
    assert 'closed' in str(error)
else:
    raise AssertionError('a closed selector answered')

left.close()
right.close()

Common Patterns¶

Dispatching on Attached Data¶

Attaching a callback as data means each ready event carries its own handler, so the loop does O(k) work on top of whatever select() costs.

import selectors
import socket

received = []

def on_readable(conn):
    received.append(conn.recv(1024))

pairs = [socket.socketpair() for _ in range(3)]

with selectors.DefaultSelector() as sel:
    for left, _ in pairs:
        left.setblocking(False)
        sel.register(left, selectors.EVENT_READ, data=on_readable)  # O(1)

    pairs[0][1].send(b'a')
    pairs[2][1].send(b'c')

    for key, events in sel.select(timeout=0):  # k = 2 ready
        key.data(key.fileobj)  # O(1) dispatch per ready file

    for left, right in pairs:
        sel.unregister(left)  # O(1) - before closing
        left.close()
        right.close()

assert sorted(received) == [b'a', b'c']

Performance Best Practices¶

✅ Do:

Use DefaultSelector; on Linux, BSD and macOS it is the O(k) implementation
unregister() a file before closing it, so the lookup stays O(1)
Keep per-connection state in data, so a ready event needs no second lookup
Close the selector, or use it in a with block, to release the kernel object

❌ Avoid:

SelectSelector or PollSelector for many mostly idle connections - every call is O(n)
SelectSelector on Unix for descriptors at or above FD_SETSIZE; select() raises ValueError

Version Notes¶

All Python 3: On Windows only sockets can be watched, because select() there accepts nothing else

select - the primitives each selector wraps
asyncio - an event loop built on a selector
socket - the file objects most often registered
socketserver - servers that wait on a selector between requests