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
asyncio Module Complexity¶
asyncio runs many coroutines on one thread. A single event loop holds two
schedules - a queue of callbacks ready now and a heap of callbacks due later -
and drains them in a loop, so nothing in this module costs what a thread would
cost, and nothing in it runs in parallel with anything else.
That shape sets what the bounds below measure. Almost every call here is O(1): it appends one entry to a schedule and returns. The costs worth knowing are the ones that are not - the calls linear in the awaitables you hand them, the heap operations that make a timer dearer than an immediate callback, the broadcasts that touch every waiter on a primitive, and the snapshots that build a fresh container of everything the loop is tracking.
No bound here prices the coroutine's own work, or the I/O it waits on. An
await that resolves after a network round trip costs a round trip; what
asyncio contributes is the scheduling around it.
Complexity Reference¶
Size variables: n = awaitables passed to one call; t = unfinished tasks alive
in the process and still registered, which is the population a task snapshot
filters; g = live async generators the loop is tracking; r = callbacks already
in the ready queue; s = timers on the loop's heap; u = timers that come due in
one iteration; x = cancelled timers sitting at the head of the heap; k = done
callbacks on one future; w = tasks waiting on one synchronization primitive; q =
items in a queue; b = bytes read or written; d = the longest separator
readuntil is given; j = how many it is given; c = frames one task's stack
exposes; f = frames and awaited tasks in a captured call graph; e = entries in
one writelines iterable; a = addresses a host name resolved to; v = argument
and environment bytes handed to a child process.
Bounds marked "+ round trip" do local work in the stated bound and then wait for a peer, the operating system, or the clock. The wait is not priced.
Running asyncio programs¶
| Operation | Time | Space | Notes |
|---|---|---|---|
asyncio.run(coro) |
O(t + g + r + s) + the coroutine | O(t + g) | Creates a loop and runs coro; everything but the coroutine is the shutdown, which cancels every unfinished task, closes the g live async generators, and clears the ready queue and the timer heap |
asyncio.Runner(...) |
O(1) | O(1) | Python 3.11+; defers loop creation to the first run() |
asyncio.Runner.run(coro) |
O(1) + the coroutine | O(t + s) | As run() without the shutdown, which close() does once for the runner's whole lifetime |
asyncio.Runner.get_loop() |
O(1) | O(1) | Creates the loop if this is the first call, then returns the stored one |
asyncio.Runner.close() |
O(t + g + r + s) | O(t + g) | Snapshots the unfinished tasks into a set and cancels them, snapshots and closes the generators, waits for the default executor, and clears both schedules |
Event loop references¶
| Operation | Time | Space | Notes |
|---|---|---|---|
asyncio.get_running_loop() |
O(1) | O(1) | Reads the thread's running-loop slot; raises RuntimeError outside a loop |
asyncio.get_event_loop() |
O(1) | O(1) | Returns the running loop inside one, or the loop set_event_loop() installed for this thread. With neither, Python 3.14 raises RuntimeError; earlier versions create one and warn, but only on the main thread and only before a loop has been set or cleared there - elsewhere they raise too |
asyncio.new_event_loop() |
O(1) | O(1) | Opens the selector and the self-pipe; the loop starts with empty schedules |
asyncio.set_event_loop(loop) |
O(1) | O(1) | Stores the loop for this thread |
asyncio.get_event_loop_policy() |
O(1) | O(1) | Deprecated in Python 3.14, for removal in 3.16 |
asyncio.set_event_loop_policy(policy) |
O(1) | O(1) | Deprecated in Python 3.14, for removal in 3.16 |
asyncio.EventLoop |
O(1) | O(1) | Python 3.13+; the platform's loop class, SelectorEventLoop on Unix |
asyncio.AbstractEventLoop |
- | - | The interface the loop classes implement; defines no bounds of its own |
AbstractEventLoop scheduling¶
The ready queue is a deque and the timer schedule is a heap, which is the whole difference between the two rows below.
| Operation | Time | Space | Notes |
|---|---|---|---|
loop.call_soon(cb, *args) |
O(1) | O(1) | Appends one Handle to the ready deque |
loop.call_soon_threadsafe(cb, *args) |
O(1) | O(1) | As call_soon, plus one byte written to the self-pipe to wake the loop |
loop.call_later(delay, cb, *args) |
O(log s) | O(1) | Pushes a TimerHandle onto the timer heap |
loop.call_at(when, cb, *args) |
O(log s) | O(1) | As call_later, with an absolute deadline on the loop's clock |
loop.time() |
O(1) | O(1) | A monotonic clock reading |
| One loop iteration | O((u + x) log s + r), plus O(s) on a sweep | O(s) while sweeping, else O(1) | Pops the u due timers off the heap and the x cancelled entries sitting at its head, then runs every callback the deque held when the iteration began. The bulk sweep replaces both: it rebuilds the heap in a new list when it holds more than 100 timers with over half cancelled, whether or not a timer is due |
loop.run_forever() |
O(1) + the callbacks | O(1) | Iterates until stop() |
loop.run_until_complete(future) |
O(1) + the future | O(1) | Adds a done callback that stops the loop, then runs it |
loop.stop() |
O(1) | O(1) | Sets a flag the current iteration checks |
loop.is_running(), loop.is_closed() |
O(1) | O(1) | Flag reads |
loop.close() |
O(r + s) | O(1) | Clears both schedules and closes the selector |
AbstractEventLoop tasks, futures and executors¶
| Operation | Time | Space | Notes |
|---|---|---|---|
loop.create_future() |
O(1) | O(1) | A bare Future bound to this loop |
loop.create_task(coro) |
O(1) | O(1) | Wraps coro in a Task and schedules its first step. An eager factory instead runs the coroutine until its first suspension before returning |
loop.set_task_factory(factory), loop.get_task_factory() |
O(1) | O(1) | Stores or reads one callable |
loop.run_in_executor(executor, fn, *args) |
O(1) + fn |
O(1) | Submits to a thread pool and wraps the concurrent.futures.Future; fn runs off the loop thread |
loop.set_default_executor(executor) |
O(1) | O(1) | Stores one executor |
loop.shutdown_default_executor() |
O(1) + the pool | O(1) | Waits for the pool's threads to finish |
loop.shutdown_asyncgens() |
O(g) | O(g) | Snapshots the live async generators and closes each |
loop.get_debug(), loop.set_debug(enabled) |
O(1) | O(1) | Flag access |
loop.set_exception_handler(handler), loop.get_exception_handler() |
O(1) | O(1) | Stores or reads one callable |
loop.call_exception_handler(context), loop.default_exception_handler(context) |
O(k log k) | O(k) | The default handler sorts the context dict's keys before formatting each into the log message |
AbstractEventLoop networking¶
Every row below does bounded local work and then waits on the network or the operating system. The local bound is what is stated; the wait is not.
| Operation | Time | Space | Notes |
|---|---|---|---|
loop.getaddrinfo(host, port), loop.getnameinfo(sockaddr) |
O(1) + round trip | O(a) | Runs the blocking resolver in the default executor |
loop.create_connection(factory, host, port) |
O(a) + round trip | O(a) | Holds the resolved address list, and the errors from the failed attempts, while trying each until one connects |
loop.create_server(factory, host, port) |
O(a) | O(a) | Binds one listening socket per resolved address |
loop.create_unix_connection(...), loop.create_unix_server(...) |
O(1) + round trip | O(1) | Unix builds only; no resolution step |
loop.create_datagram_endpoint(factory, ...) |
O(a) + round trip | O(a) | Resolves the local and remote names, pairs them into a candidate list, and tries each until one socket binds |
loop.connect_accepted_socket(factory, sock) |
O(1) | O(1) | Wraps an already-accepted socket in a transport |
loop.connect_read_pipe(factory, pipe), loop.connect_write_pipe(factory, pipe) |
O(1) | O(1) | Registers one pipe with the selector |
loop.start_tls(transport, protocol, context) |
O(1) + round trip | O(1) | Handshake over an open transport |
loop.sendfile(transport, file), loop.sock_sendfile(sock, file) |
O(b) + round trip | O(1) | Uses the kernel's sendfile where available, so the bytes need not pass through Python |
loop.add_signal_handler(sig, cb), loop.remove_signal_handler(sig) |
O(1) | O(1) | Unix builds only |
Handle and TimerHandle¶
| Operation | Time | Space | Notes |
|---|---|---|---|
Handle.cancel() |
O(1) | O(1) | Clears the callback; the loop skips the entry when it reaches it, rather than removing it now |
Handle.cancelled() |
O(1) | O(1) | Flag read |
Handle.get_context() |
O(1) | O(1) | Python 3.12+; the contextvars.Context captured at scheduling time |
TimerHandle.when() |
O(1) | O(1) | The stored deadline |
TimerHandle.cancel() |
O(1) | O(1) | As Handle.cancel(), and increments the loop's cancelled-timer count |
Future¶
| Operation | Time | Space | Notes |
|---|---|---|---|
asyncio.Future(loop=...) |
O(1) | O(1) | Starts pending with no callbacks |
Future.result(), Future.exception() |
O(1) | O(1) | Returns the stored object; raises InvalidStateError while pending |
Future.set_result(value), Future.set_exception(exc) |
O(k) | O(k) | Each of the k done callbacks is scheduled with call_soon, so the ready queue grows by k handles rather than the callbacks running here |
Future.done(), Future.cancelled() |
O(1) | O(1) | State reads |
Future.cancel(msg=None) |
O(k) | O(k) | Marks cancelled and schedules the same k callbacks |
Future.add_done_callback(cb) |
O(1) | O(1) | Appends; on an already-done future, schedules cb instead |
Future.remove_done_callback(cb) |
O(k) | O(k) | Rebuilds the callback list without cb |
Future.get_loop() |
O(1) | O(1) | The stored loop |
await future |
O(1) | O(1) | Suspends the awaiting task and registers one done callback |
Task¶
Task is a Future, with two exceptions: set_result() and
set_exception() raise RuntimeError, because a task's result comes from its
coroutine. Every other Future row applies. These are the additions.
| Operation | Time | Space | Notes |
|---|---|---|---|
asyncio.Task(coro, loop=...) |
O(1) | O(1) | Registers the task with the loop and schedules the first step |
Task.get_coro(), Task.get_name() |
O(1) | O(1) | Stored attributes |
Task.get_context() |
O(1) | O(1) | Python 3.12+; the contextvars.Context the task runs its steps in |
Task.set_name(value) |
O(1) | O(1) | str(value) is stored |
Task.cancel(msg=None) |
O(1) plus the awaited future's own cancel() |
O(1) | Arranges for CancelledError at the task's next step, or cancels the future it is currently waiting on. That future's cancel is not always O(1): a task suspended on gather() cancels each of its n children here, synchronously. The coroutine still handles the error later |
Task.cancelling(), Task.uncancel() |
O(1) | O(1) | Python 3.11+; read and decrement the cancellation counter TaskGroup and timeout() use |
Task.get_stack(limit=None) |
O(c) | O(c) | A suspended task exposes the single frame it is suspended in, however deep the coroutines below it; a task that raised exposes its traceback's frames |
Task.print_stack(limit=None, file=None) |
O(c) | O(c) | Formats the same frames |
Task.result(), Task.exception(), Task.cancelled() |
O(1) | O(1) | Future's, unchanged |
Task.add_done_callback(cb), Task.remove_done_callback(cb) |
O(1), O(k) | O(1), O(k) | Future's, unchanged |
Extending Task and Future¶
These four hooks let an alternative Task implementation keep all_tasks()
and current_task() accurate. They are documented for that purpose and are not
part of ordinary use.
| Operation | Time | Space | Notes |
|---|---|---|---|
asyncio._register_task(task) |
O(1) | O(1) | Adds the task to the set all_tasks() snapshots |
asyncio._unregister_task(task) |
O(1) | O(1) | Removes it |
asyncio._enter_task(loop, task) |
O(1) | O(1) | Sets the loop's current-task slot; raises RuntimeError if one is already running |
asyncio._leave_task(loop, task) |
O(1) | O(1) | Clears it, raising RuntimeError if task is not the one entered |
Coroutine and task functions¶
| Operation | Time | Space | Notes |
|---|---|---|---|
asyncio.create_task(coro, name=None) |
O(1) | O(1) | Requires a running loop; delegates to loop.create_task |
asyncio.ensure_future(obj) |
O(1) | O(1) | Returns a Future unchanged, wraps a coroutine in a Task, and wraps any other awaitable in a task around it |
asyncio.current_task(loop=None) |
O(1) | O(1) | Reads the loop's current-task slot |
asyncio.all_tasks(loop=None) |
O(t) | O(t) | Scans the process-wide task registry, not just this loop's tasks, and builds a new set of the ones belonging to loop on every call. Before Python 3.14 a finished task stays registered while anything still references it, so keeping completed tasks around makes every call dearer |
asyncio.sleep(0) |
O(1) | O(1) | Yields to the loop once without scheduling a timer |
asyncio.sleep(delay) |
O(log s) | O(1) | One timer on the heap for a positive delay |
asyncio.iscoroutine(obj), asyncio.isfuture(obj) |
O(1) | O(1) | Type checks |
asyncio.iscoroutinefunction(fn) |
O(1) | O(1) | Deprecated in Python 3.14, for removal in 3.16; use inspect.iscoroutinefunction |
asyncio.eager_task_factory(loop, coro, ...) |
O(1) + the coroutine's first step | O(1) | Python 3.12+; runs the coroutine until it first suspends, and returns a finished task if it never does |
asyncio.create_eager_task_factory(task_class) |
O(1) | O(1) | Python 3.12+; builds such a factory for a Task subclass |
Running awaitables concurrently¶
| Operation | Time | Space | Notes |
|---|---|---|---|
asyncio.gather(*aws) |
O(n) setup, O(n) on completion | O(n) | Wraps each awaitable and adds a done callback; assembles the results list in argument order once all n finish |
asyncio.wait(aws, timeout=None, return_when=...) |
O(n) | O(n) | Adds a callback to each, then partitions them into the done and pending sets. A timeout adds one timer. From Python 3.11 the awaitables must already be tasks or futures; a bare coroutine raises TypeError |
asyncio.as_completed(aws, timeout=None) |
O(n) setup | O(n) | Yields n results in completion order. Python 3.13+ also supports async for, which hands back each input that was already a task or future, and the wrapping task for an input that was a bare coroutine; earlier versions yield wrapper coroutines that must be awaited for the result |
asyncio.TaskGroup() |
O(1) | O(1) | Python 3.11+ |
TaskGroup.create_task(coro) |
O(1) | O(1) | Adds the task to the group's set |
async with TaskGroup() exit |
O(n) | O(n) | n = tasks created in the group; waits for all, and on failure cancels the rest and collects an ExceptionGroup |
asyncio.shield(aw) |
O(1) | O(1) | One outer future; cancelling it leaves the inner awaitable running |
asyncio.wait_for(aw, timeout) |
O(log s) | O(1) | One timer around one awaitable |
asyncio.timeout(delay), asyncio.timeout_at(when) |
O(1) | O(1) | Python 3.11+; the context manager object. Entering it schedules one timer |
Timeout.when(), Timeout.expired() |
O(1) | O(1) | Stored deadline and flag |
Timeout.reschedule(when) |
O(log s) | O(1) | Cancels the pending timer and pushes a new one |
asyncio.to_thread(fn, *args) |
O(1) + fn |
O(1) | Runs fn on the default thread pool with the caller's context copied |
asyncio.run_coroutine_threadsafe(coro, loop) |
O(1) | O(1) | Schedules the coroutine on another thread's loop; returns a concurrent.futures.Future |
asyncio.wrap_future(future, loop=None) |
O(1) | O(1) | Bridges a concurrent.futures.Future to an asyncio.Future |
Call graph introspection¶
Python 3.14+. capture_call_graph() walks the awaited-by links a task records.
| Operation | Time | Space | Notes |
|---|---|---|---|
asyncio.capture_call_graph(future=None, depth=0, limit=None) |
O(f) | O(f) | Walks the frames of the target and the tasks awaiting it |
asyncio.format_call_graph(...), asyncio.print_call_graph(...) |
O(f·h) | O(f·h) | h = the depth of the awaited-by chain; each line is indented one level per step, so a deep chain renders more text than a wide graph with the same f |
asyncio.future_add_to_awaited_by(fut, waiter) |
O(1) | O(1) | Records one awaiter on the future |
asyncio.future_discard_from_awaited_by(fut, waiter) |
O(1) | O(1) | Removes it |
asyncio.FutureCallGraph, asyncio.FrameCallGraphEntry |
O(1) | O(1) | The named tuples a capture is built from |
Queue, LifoQueue and PriorityQueue¶
The three differ only in the container behind them: a deque, a list used as a stack, and a heap.
| Operation | Time | Space | Notes |
|---|---|---|---|
asyncio.Queue(maxsize=0) |
O(1) | O(1) | Unbounded by default |
Queue.put_nowait(item), Queue.get_nowait() |
O(1) | O(1) | Raises QueueFull or QueueEmpty rather than waiting. PriorityQueue is O(log q) for both |
Queue.put(item) |
O(1) | O(1) | Waits only when the queue is full; PriorityQueue is O(log q) |
Queue.get() |
O(1) | O(1) | Waits only when the queue is empty; PriorityQueue is O(log q) |
Queue.qsize(), Queue.empty(), Queue.full() |
O(1) | O(1) | Length and bound comparisons |
Queue.maxsize |
O(1) | O(1) | The stored limit; 0 means unbounded |
Queue.task_done() |
O(1), O(w) at zero | O(1), O(w) at zero | Decrements the unfinished count; the call that brings it to zero wakes every join() waiter, queueing w callbacks |
Queue.join() |
O(1) | O(1) | Returns at once when nothing is unfinished |
Queue.shutdown(immediate=False) |
O(w); immediate adds O(q), or O(q log q) for PriorityQueue |
O(w) | Python 3.13+; wakes every waiter blocked in get() or put() so each raises QueueShutDown. With immediate, also drains the q queued items one at a time through the queue's own get - a heap pop for PriorityQueue - and releases the join() waiters; without it they keep waiting until task_done() balances the outstanding work |
asyncio.LifoQueue, asyncio.PriorityQueue |
as above | O(q) | Last-in-first-out over a list, and lowest-first over a heap |
Lock, Semaphore, BoundedSemaphore, Event, Condition and Barrier¶
Each primitive keeps a deque of waiter futures. Acquiring and releasing touch one; the broadcasts touch all w. A waiter that gives up removes its own future from that deque, which is O(1) at either end and O(w) from the middle - so a cancelled waiter deep in a long queue is the one case where an otherwise constant-time wait is not.
| Operation | Time | Space | Notes |
|---|---|---|---|
Lock.acquire() |
O(1), O(w) past cancelled waiters | O(1) | Uncontended with no queue, sets a flag. Contended, appends one future and waits. The uncontended path scans the queue when one exists, to check that every waiter in it was cancelled |
Lock.release() |
O(1) | O(1) | Clears the flag and wakes the first queued waiter, unless that waiter is already done |
Lock.locked() |
O(1) | O(1) | Flag read |
Semaphore(value), Semaphore.acquire(), Semaphore.release() |
O(1), O(w) past done waiters | O(1) | A counter and the same waiter deque. release() scans from the front for the first waiter that is not already done |
Semaphore.locked() |
O(w) | O(1) | True when the counter is exhausted, and also while any queued waiter is uncancelled, which FIFO order requires it to scan for |
BoundedSemaphore(value) |
O(1) | O(1) | As Semaphore, but release() past the initial value raises ValueError |
Event.set() |
O(w) | O(w) | Resolves every waiter's future, which queues w callbacks on the loop |
Event.clear(), Event.is_set() |
O(1) | O(1) | Flag access |
Event.wait() |
O(1) | O(1) | Returns immediately when the flag is set |
Condition(lock=None) |
O(1) | O(1) | Wraps a Lock |
Condition.wait() |
O(1) | O(1) | Releases the lock, waits, then reacquires it |
Condition.wait_for(predicate) |
O(p) per wake | O(1) | p = the predicate's own cost; re-evaluated after each notification until it is true |
Condition.notify(n=1) |
O(w) | O(n) | Walks the waiter deque until n of them have been woken, skipping any already done, so it is O(n) only when the front of the queue is live |
Condition.notify_all() |
O(w) | O(w) | notify(len(waiters)) |
Barrier(parties) |
O(1) | O(1) | Python 3.11+ |
Barrier.wait() |
O(1), O(w) twice a cycle | O(1), O(w) twice a cycle | The arrival that fills the barrier wakes every participant, and the last of them to leave wakes whoever is queued for the next cycle. Both are broadcasts; every other arrival and departure is constant |
Barrier.parties, Barrier.n_waiting, Barrier.broken |
O(1) | O(1) | Counter and flag reads |
Barrier.reset(), Barrier.abort() |
O(w) | O(w) | Both wake every waiter; abort() leaves the barrier broken so later arrivals raise BrokenBarrierError |
Streams¶
| Operation | Time | Space | Notes |
|---|---|---|---|
asyncio.open_connection(host, port, limit=65536) |
O(a) + round trip | O(limit) | a = resolved addresses; the reader's buffer grows to limit before readuntil refuses |
asyncio.open_unix_connection(path, ...) |
O(1) + round trip | O(limit) | Unix builds only |
asyncio.start_server(cb, host, port, ...) |
O(a) | O(a) | One listening socket per resolved address |
asyncio.start_unix_server(cb, path, ...) |
O(1) | O(1) | Unix builds only |
StreamReader and StreamWriter¶
| Operation | Time | Space | Notes |
|---|---|---|---|
StreamReader.read(n=-1) |
O(b) | O(b) | n = -1 reads to EOF and holds the whole stream in memory; a positive n returns at most n bytes |
StreamReader.readline() |
O(b + d) | O(b) | readuntil(b'\n'), but returns the partial line at EOF rather than raising IncompleteReadError. Over the stream's limit it raises ValueError, and unlike readuntil it drops the oversized data from the buffer |
StreamReader.readuntil(separator=b'\n') |
O(j log j + j(b + d)) | O(b + j) | j = separators, 1 unless a tuple is passed, which Python 3.13+ is what accepts at all; they are sorted by length first, so the shortest match wins. The buffer is scanned once per separator. The scan resumes from a remembered offset, re-examining only the last d - 1 bytes so a separator straddling two chunks is still found. Raises LimitOverrunError past the stream's limit and IncompleteReadError at EOF |
StreamReader.readexactly(n) |
O(n) | O(n) | Raises IncompleteReadError if EOF arrives first |
StreamReader.at_eof() |
O(1) | O(1) | True once the buffer is empty and EOF was fed |
StreamReader.feed_eof() |
O(1) | O(1) | Marks the end and wakes a waiting read |
async for line in reader |
O(b + d) per line | O(b) | readline() until it returns empty |
StreamWriter.write(data) |
O(b + e) | O(b) | Buffers on the transport without waiting for the socket, then checks whether to pause writing - which sums the e chunks already buffered |
StreamWriter.writelines(data) |
O(b + e) | O(b) | e = items in the iterable. The transport's default implementation concatenates them and writes the result; the socket transport overrides it to buffer one memoryview per item instead, making it O(e) with no copy of the bytes |
StreamWriter.drain() |
O(1) + round trip | O(1) | Returns at once unless the transport's buffer is over its high-water mark |
StreamWriter.write_eof(), StreamWriter.can_write_eof() |
O(1) | O(1) | Half-close support |
StreamWriter.close(), StreamWriter.is_closing() |
O(1) | O(1) | Flushes what is buffered, then closes |
StreamWriter.wait_closed() |
O(1) + the flush | O(1) | Waits for the transport to finish closing |
StreamWriter.transport |
O(1) | O(1) | The stored transport |
StreamWriter.get_extra_info(name, default=None) |
O(1) | O(1) | One dict lookup on the transport |
StreamWriter.start_tls(context) |
O(1) + round trip | O(1) | Python 3.11+ |
asyncio.StreamReaderProtocol |
O(1) | O(1) | The protocol wiring a reader and writer to a transport |
Server¶
| Operation | Time | Space | Notes |
|---|---|---|---|
Server.sockets |
O(k) | O(k) | k = listening sockets; a fresh tuple each call |
Server.is_serving() |
O(1) | O(1) | Flag read |
Server.start_serving(), Server.serve_forever() |
O(k) | O(k) | Registers an accept callback per listening socket; serve_forever() then waits |
Server.close() |
O(k) | O(1) | Stops accepting; open connections continue |
Server.close_clients(), Server.abort_clients() |
O(m) | O(m) | Python 3.13+; m = open connections; both copy the connection set before closing them gracefully or at once |
Server.wait_closed() |
O(1) + the connections | O(1) | Waits for the sockets and the remaining handlers |
Server.get_loop() |
O(1) | O(1) | The stored loop |
asyncio.AbstractServer |
- | - | The interface Server implements |
Subprocesses¶
| Operation | Time | Space | Notes |
|---|---|---|---|
asyncio.create_subprocess_exec(program, *args) |
O(v) + the spawn | O(v) | v = the argument list and the environment passed to the child, both copied on the way to fork and exec |
asyncio.create_subprocess_shell(cmd) |
O(v) + the spawn | O(v) | v is the command string and the environment; a shell is spawned to parse cmd |
Process.pid, Process.returncode |
O(1) | O(1) | Stored values; returncode is None until exit |
Process.stdin, Process.stdout, Process.stderr |
O(1) | O(1) | The stream objects, or None for a pipe that was not requested |
Process.wait() |
O(1) + the child | O(1) | Waits for exit. Deadlocks if the child fills a pipe nobody is draining |
Process.communicate(input=None) |
O(b) | O(b) | Writes input, drains both output pipes concurrently, and returns them whole |
Process.send_signal(sig), Process.terminate(), Process.kill() |
O(1) | O(1) | One signal to the child |
Transports¶
| Operation | Time | Space | Notes |
|---|---|---|---|
BaseTransport.close(), BaseTransport.is_closing() |
O(1) | O(1) | Flushes what is buffered, then closes |
BaseTransport.get_extra_info(name, default=None) |
O(1) | O(1) | One dict lookup |
BaseTransport.get_protocol(), BaseTransport.set_protocol(p) |
O(1) | O(1) | Attribute access |
ReadTransport.is_reading(), ReadTransport.pause_reading(), ReadTransport.resume_reading() |
O(1) | O(1) | Adds or removes the socket's read registration |
WriteTransport.write(data) |
O(b + e) | O(b) | Writes what the socket accepts now, buffers the rest, then runs the flow-control check over the e buffered chunks |
WriteTransport.writelines(data) |
O(b + e) | O(b) | The base implementation joins the items and calls write(). A transport may override it: the socket one buffers a memoryview per item and writes them vectored, which avoids the copy |
WriteTransport.write_eof(), WriteTransport.can_write_eof() |
O(1) | O(1) | Half-close |
WriteTransport.abort() |
O(1) | O(1) | Closes at once and discards the buffer |
WriteTransport.get_write_buffer_size() |
O(e) | O(1) | e = chunks currently buffered; the socket transport sums their lengths rather than keeping a running total, and the flow-control check that decides whether to pause writing calls it on every write |
WriteTransport.get_write_buffer_limits() |
O(1) | O(1) | The stored watermarks behind pause_writing and drain() |
WriteTransport.set_write_buffer_limits(high, low) |
O(e) | O(1) | Stores the watermarks and re-runs the flow-control check against them |
DatagramTransport.sendto(data, addr=None) |
O(b) | O(b) | One datagram |
DatagramTransport.abort() |
O(1) | O(1) | Closes without flushing |
SubprocessTransport.get_pid(), SubprocessTransport.get_returncode() |
O(1) | O(1) | Stored values |
SubprocessTransport.get_pipe_transport(fd) |
O(1) | O(1) | One dict lookup |
SubprocessTransport.send_signal(sig), SubprocessTransport.terminate(), SubprocessTransport.kill() |
O(1) | O(1) | One signal |
asyncio.Transport |
- | - | ReadTransport and WriteTransport together |
Protocols¶
Protocol methods are callbacks the loop invokes; their cost is the
application's. The bounds below are what asyncio does around each call.
| Operation | Time | Space | Notes |
|---|---|---|---|
BaseProtocol.connection_made(transport), BaseProtocol.connection_lost(exc) |
O(1) | O(1) | Called once each |
BaseProtocol.pause_writing(), BaseProtocol.resume_writing() |
O(1) | O(1) | Called as the write buffer crosses its watermarks |
Protocol.data_received(data) |
O(1) | O(b) | Called with each chunk the socket produced; asyncio does not reassemble them |
Protocol.eof_received() |
O(1) | O(1) | Return true to keep the transport open |
BufferedProtocol.get_buffer(sizehint), BufferedProtocol.buffer_updated(nbytes), BufferedProtocol.eof_received() |
O(1) | O(1) | The socket reads into the protocol's own buffer, so no per-chunk bytes object is allocated |
DatagramProtocol.datagram_received(data, addr), DatagramProtocol.error_received(exc) |
O(1) | O(b) | One call per datagram |
SubprocessProtocol.pipe_data_received(fd, data), SubprocessProtocol.pipe_connection_lost(fd, exc), SubprocessProtocol.process_exited() |
O(1) | O(b) | One call per chunk or event |
Exceptions¶
| Operation | Time | Space | Notes |
|---|---|---|---|
asyncio.CancelledError |
O(1) | O(1) | Inherits BaseException, not Exception, so except Exception does not swallow a cancellation |
asyncio.TimeoutError |
O(1) | O(1) | The builtin TimeoutError from Python 3.11 on; a distinct class before that |
asyncio.InvalidStateError |
O(1) | O(1) | Result or exception set on a done future, or read from a pending one |
asyncio.IncompleteReadError |
O(1) | O(1) | IncompleteReadError.partial holds the bytes read, IncompleteReadError.expected the number wanted |
asyncio.LimitOverrunError |
O(1) | O(1) | LimitOverrunError.consumed says how far the scan got; the data stays in the buffer |
asyncio.QueueEmpty, asyncio.QueueFull, asyncio.QueueShutDown |
O(1) | O(1) | QueueShutDown is Python 3.13+ |
asyncio.BrokenBarrierError |
O(1) | O(1) | Python 3.11+ |
asyncio.SendfileNotAvailableError |
O(1) | O(1) | The platform has no usable sendfile for this pair |
Two Schedules, Two Costs¶
The loop holds ready callbacks in a deque and timers in a heap. An immediate
callback is an append; a delayed one is a heap push. sleep(0) is the special
case: it yields to the loop without becoming a timer at all.
import asyncio
import asyncio.base_events
import heapq
pushes = 0
real_heappush = heapq.heappush
def counting_heappush(heap, item):
global pushes
pushes += 1
return real_heappush(heap, item)
async def main():
global pushes
loop = asyncio.get_running_loop()
asyncio.base_events.heapq.heappush = counting_heappush
try:
# O(1): straight onto the ready deque
pushes = 0
loop.call_soon(lambda: None)
assert pushes == 0
# O(log s): onto the timer heap
pushes = 0
handle = loop.call_later(10, lambda: None)
assert pushes == 1
handle.cancel()
# sleep(0) yields without scheduling a timer
pushes = 0
await asyncio.sleep(0)
assert pushes == 0
# any positive delay does
pushes = 0
await asyncio.sleep(0.001)
assert pushes == 1
finally:
asyncio.base_events.heapq.heappush = real_heappush
asyncio.run(main())
Cancelling a timer does not remove it from the heap: TimerHandle.cancel()
clears the callback and leaves the entry in place. The loop drops the dead
entries in two ways, and which one applies decides how long they are carried.
Cancelled entries at the head are popped on the next iteration, so cancelling
the soonest timeout reclaims it promptly. An entry behind a live one is
carried until either it reaches the head in its turn or the bulk sweep runs,
and that sweep needs the heap to hold more than 100 timers with over half of
them cancelled. A loop that keeps a long-lived timer in front of many cancelled
ones therefore holds on to them.
gather and wait Are Linear in Their Awaitables¶
gather() wraps every awaitable before the first one runs, so its setup grows
with n even though each individual wrap is O(1), and it then assembles a result
list in argument order. wait() does not wrap: from Python 3.11 it requires
tasks or futures and refuses a bare coroutine. Its n is the callback it adds to
each, and the two sets it partitions them into.
import asyncio
async def value(x):
return x
async def main():
# gather: n wraps up front, one list of n results in argument order
results = await asyncio.gather(*[value(i) for i in range(5)])
assert results == [0, 1, 2, 3, 4]
# wait: one callback added per task, two sets out
tasks = [asyncio.create_task(value(i)) for i in range(5)]
done, pending = await asyncio.wait(tasks)
assert len(done) == 5
assert pending == set()
# FIRST_COMPLETED returns as soon as one finishes - the rest stay pending
slow = [asyncio.create_task(asyncio.sleep(10)) for _ in range(3)]
quick = asyncio.create_task(value(1))
done, pending = await asyncio.wait(
[*slow, quick], return_when=asyncio.FIRST_COMPLETED
)
assert quick in done
assert len(pending) == 3
for task in pending:
task.cancel()
await asyncio.gather(*pending, return_exceptions=True)
asyncio.run(main())
asyncio.TaskGroup costs the same n, charged one task at a time: each
create_task() is O(1), and the async with exit waits for whatever the group
accumulated. Unlike gather(), a failure there cancels the siblings and raises
an ExceptionGroup.
all_tasks Builds a Fresh Set¶
current_task() reads one slot. all_tasks() filters every task the loop is
tracking and returns a new set, so it is O(t) in both time and space each
time it is called - it is a snapshot, not a view.
import asyncio
async def main():
tasks = [asyncio.create_task(asyncio.sleep(10)) for _ in range(4)]
await asyncio.sleep(0)
first = asyncio.all_tasks()
second = asyncio.all_tasks()
# Equal contents, but a new set each call - O(t), not a stored one
assert first == second
assert first is not second
# The caller's own task is in the snapshot; current_task() is the O(1) way
assert asyncio.current_task() in first
assert len(first) == len(tasks) + 1
for task in tasks:
task.cancel()
await asyncio.gather(*tasks, return_exceptions=True)
asyncio.run(main())
A Broadcast Touches Every Waiter¶
Acquiring and releasing a lock or semaphore touches one waiter. Event.set(),
Condition.notify_all() and a barrier's final arrival touch all w of them, so
those are the rows where the waiter count reaches the bound.
import asyncio
async def main():
event = asyncio.Event()
waiters = [asyncio.create_task(event.wait()) for _ in range(8)]
await asyncio.sleep(0)
# All eight are queued on the one event
assert len(event._waiters) == 8
# O(w): set() resolves every one of them
event.set()
assert await asyncio.gather(*waiters) == [True] * 8
# Once set, waiting is O(1) and never queues
assert event.is_set()
await event.wait()
assert not event._waiters
condition = asyncio.Condition()
woken = []
waiting = [asyncio.create_task(_wait_on(condition, woken, i)) for i in range(5)]
while len(condition._waiters) < 5:
await asyncio.sleep(0)
# notify(2) wakes exactly two, however many are queued
async with condition:
condition.notify(2)
for _ in range(8):
await asyncio.sleep(0)
assert len(woken) == 2
# notify_all() wakes whatever is left
async with condition:
condition.notify_all()
await asyncio.gather(*waiting)
assert len(woken) == 5
async def _wait_on(condition, woken, index):
async with condition:
await condition.wait()
woken.append(index)
asyncio.run(main())
A Queue's Cost Is Its Container¶
Queue is a deque and LifoQueue is a list used as a stack, so both move an
item in O(1). PriorityQueue is a heap, which buys ordering for O(log q).
import asyncio
async def main():
fifo = asyncio.Queue()
lifo = asyncio.LifoQueue()
prio = asyncio.PriorityQueue()
for i in (3, 1, 2):
fifo.put_nowait(i) # O(1) deque append
lifo.put_nowait(i) # O(1) list append
prio.put_nowait(i) # O(log q) heap push
assert [fifo.get_nowait() for _ in range(3)] == [3, 1, 2]
assert [lifo.get_nowait() for _ in range(3)] == [2, 1, 3]
assert [prio.get_nowait() for _ in range(3)] == [1, 2, 3]
# qsize() is a length, not a walk
bounded = asyncio.Queue(maxsize=2)
await bounded.put("a")
await bounded.put("b")
assert bounded.qsize() == 2 and bounded.full()
# A full queue refuses rather than growing
try:
bounded.put_nowait("c")
except asyncio.QueueFull:
pass
else:
raise AssertionError("expected QueueFull")
# join() returns at once when nothing is outstanding
bounded.get_nowait()
bounded.get_nowait()
bounded.task_done()
bounded.task_done()
await bounded.join()
asyncio.run(main())
An Eager Task Runs Before create_task Returns¶
The default factory schedules a task's first step and returns; the coroutine
body does not run until the loop next iterates. An eager factory runs the body
until its first suspension inside create_task(), which skips a loop iteration
for coroutines that never suspend.
import asyncio
import sys
async def main():
loop = asyncio.get_running_loop()
order = []
async def record(tag):
order.append(tag)
# Default: the body runs after create_task() returns
task = asyncio.create_task(record("body"))
order.append("after create_task")
await task
assert order == ["after create_task", "body"]
if sys.version_info < (3, 12):
return
# Eager: the body runs during create_task()
order.clear()
loop.set_task_factory(asyncio.eager_task_factory)
task = asyncio.create_task(record("body"))
order.append("after create_task")
assert order == ["body", "after create_task"]
assert task.done() # never suspended, so it finished eagerly
await task
loop.set_task_factory(None)
asyncio.run(main())
Cancellation Is a Request¶
Task.cancel() does not run the cancellation: it arranges for
CancelledError to be raised at the task's next step, and the task decides
what happens then. The call itself is O(1), but it forwards
to whatever future the task is suspended on, and that future's cancel() is
charged here: a plain future schedules its k done callbacks, and a gather()
cancels each of its n children there and then. shield()
lets an outer cancellation resolve without reaching the inner work at all.
import asyncio
async def main():
async def work():
try:
await asyncio.sleep(0.05)
return "finished"
except asyncio.CancelledError:
return "caught"
inner = asyncio.create_task(work())
outer = asyncio.shield(inner)
await asyncio.sleep(0)
# Cancelling the shield does not cancel what it wraps
outer.cancel()
try:
await outer
except asyncio.CancelledError:
pass
else:
raise AssertionError("expected the shield to cancel")
assert await inner == "finished"
assert not inner.cancelled()
# CancelledError is not an Exception, so a bare except Exception misses it
assert issubclass(asyncio.CancelledError, BaseException)
assert not issubclass(asyncio.CancelledError, Exception)
asyncio.run(main())
asyncio Is for Waiting, Not for Computing¶
One loop runs on one thread, so a coroutine that computes without awaiting blocks every other task until it returns. Concurrency here comes from tasks being suspended, which helps when they are waiting on something external and does nothing when they are not.
Threads are not the answer to that either: CPython's global interpreter lock
lets one thread at a time execute bytecode, so moving pure-Python computation
onto threads leaves the same single core doing the same work. Use
multiprocessing or concurrent.futures.ProcessPoolExecutor for CPU-bound
work; use asyncio.to_thread() to keep the loop responsive across a blocking
call, which releases the lock while it waits.
import asyncio
import time
def blocking_call():
"""Stands in for a library that has no async form."""
time.sleep(0.05)
return "done"
async def main():
# One call, to measure what three of them would cost in sequence
start = time.perf_counter()
await asyncio.to_thread(blocking_call)
one = time.perf_counter() - start
# The same three waits, overlapped on the default thread pool
start = time.perf_counter()
results = await asyncio.gather(
asyncio.to_thread(blocking_call),
asyncio.to_thread(blocking_call),
asyncio.to_thread(blocking_call),
)
three = time.perf_counter() - start
assert results == ["done", "done", "done"]
# Overlapped, not summed: nothing like 3x one call
assert three < 2 * one
asyncio.run(main())
Version Notes¶
- Python 3.11+:
TaskGroup,timeout(),timeout_at(),Runner,BarrierandBrokenBarrierError;Task.cancelling()andTask.uncancel();asyncio.TimeoutErrorbecomes the builtinTimeoutError;StreamWriter.start_tls() - Python 3.12+:
eager_task_factoryandcreate_eager_task_factory();Handle.get_context()andTask.get_context(); the child watcher API is deprecated - Python 3.13+:
asyncio.EventLoop;Queue.shutdown()andQueueShutDown;as_completed()supportsasync for, handing back each task or future it was given;readuntil()accepts a tuple of separators;Server.close_clients()andServer.abort_clients() - Python 3.14+:
capture_call_graph(),format_call_graph(),print_call_graph(),future_add_to_awaited_by()andfuture_discard_from_awaited_by();get_event_loop()raisesRuntimeErrorwhen no loop is running and none was set for the thread, instead of creating one - a loop installed withset_event_loop()is still returned; a finished task is dropped from the registryall_tasks()scans, where earlier versions keep it while it is referenced; the child watcher API andget_child_watcher()are removed;iscoroutinefunction(),get_event_loop_policy()andset_event_loop_policy()are deprecated for removal in 3.16
Related Documentation¶
- Threading Module - The same primitives with real preemption, and the lock that limits them
- Concurrent.futures Module - The pools behind
to_thread()andrun_in_executor() - Multiprocessing Module - Where CPU-bound work belongs
- Queue Module - The blocking queues
asyncio.Queuemirrors - Heapq Module - The heap under the timer schedule and
PriorityQueue - Socket Module - What the transports are built on
- Contextvars Module - The context a task and a
Handlecapture