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
sys Module Complexity¶
The sys module is the interpreter talking about itself: the module table, the import path, the
standard streams, the exception currently being handled, the tracing and auditing hooks, and a
long list of values fixed when the process started. Almost every name on it reads or writes one
field, which is why almost every row below is O(1).
The two big containers are not copies: sys.modules and sys.path are the interpreter's own
dict and list, so what you mutate is what the import system uses next. The rows worth knowing
are the handful that are not constant: walking the frame chain, visiting every live thread,
interning a string, running the audit hooks, and instrumenting code for sys.monitoring.
d is the frame depth a walk covers, t the live threads, h the installed audit hooks, m
the entries in sys.modules, e the entries in sys.path, b the entries in
sys.builtin_module_names, f the frames on all thread stacks at once, c the bytecode of one
code object, C the bytecode of all the distinct code objects on those stacks, r the
characters in a value's repr, and i the live interpreters. Dict and set bounds treat key
hashing as O(1).
Complexity Reference¶
Exception state and exit¶
| Operation | Time | Space | Notes |
|---|---|---|---|
sys.exc_info() |
O(1) | O(1) | Builds a 3-tuple from the thread's current exception |
sys.exception() |
O(1) | O(1) | The exception itself, with no tuple built; Python 3.11+ |
sys.last_exc, sys.last_type, sys.last_value, sys.last_traceback |
O(1) | O(1) | Bound by the interpreter when it reports an unhandled exception, just before it calls sys.excepthook - so a replacement hook and anything running after it can read them, and calling the hook yourself does not bind them. Before that, reading one raises AttributeError. last_exc is Python 3.12+ |
sys.excepthook(type, value, traceback), sys.__excepthook__ |
O(d + s + m) | O(s + m) | Walks the whole traceback and reads the source line for each frame it prints; s = characters of source read, m = characters the exception itself renders to - its str() plus any __notes__. A repeated frame prints three times and then a count, and a chained exception adds its own frames |
sys.unraisablehook(unraisable), sys.__unraisablehook__ |
O(d + s + m) | O(s + m) | The same formatting for an exception that had nowhere to propagate, such as one raised in __del__; here m also covers the repr() of the object the report names |
sys.tracebacklimit |
O(1) | O(1) | Caps how many frames the two rows above print, not how many they walk; unbound by default, which means no cap |
sys.exit(code) |
O(1) | O(1) | Raises SystemExit, so this is the cost of the raise, not of the shutdown that follows |
Frames and threads¶
| Operation | Time | Space | Notes |
|---|---|---|---|
sys._getframe(depth=0) |
O(d) | O(1) | Follows the frame chain one link at a time; ValueError past the bottom |
sys._getframemodulename(depth=0) |
O(d) | O(1) | The same walk, returning __name__ from that frame's globals; Python 3.12+ |
sys._current_frames() |
O(t) | O(t) | One entry per live thread |
sys._current_exceptions() |
O(t) | O(t) | One entry per live thread as well: a thread handling nothing maps to None rather than being left out. Python 3.12+ stores the exception; earlier versions store an exc_info() triple |
sys.getrecursionlimit() |
O(1) | O(1) | Reads one interpreter field |
sys.setrecursionlimit(limit) |
O(t) | O(1) | Writes through to every thread state from Python 3.11; a limit below the current depth raises RecursionError |
sys.getswitchinterval(), sys.setswitchinterval(interval) |
O(1) | O(1) | One field each |
Tracing, profiling and auditing¶
| Operation | Time | Space | Notes |
|---|---|---|---|
sys.settrace(function), sys.setprofile(function) |
O(f + C) | O(C) | From Python 3.12 these run on the sys.monitoring machinery, so installing one pauses every thread and instruments the code objects on their stacks; on 3.10 and 3.11 it is O(1). Either way the bill that matters comes afterwards, one call per traced event |
sys.gettrace(), sys.getprofile() |
O(1) | O(1) | Reads the installed function back |
sys.call_tracing(func, args) |
O(1) plus func |
O(1) | Saves and restores the tracing state around one call, so a trace function can re-enter itself; on Python 3.11 the call runs untraced instead |
sys.audit(event, *args) |
O(h) | O(1) | Hooks are called in the order they were added, until one of them raises - that exception propagates and the rest are skipped |
sys.addaudithook(hook) |
O(h) | O(1) | Raises the sys.addaudithook event first, so every hook already installed runs before the append; there is no matching remove |
sys.activate_stack_trampoline(backend), sys.deactivate_stack_trampoline() |
O(1) | O(1) | Switches the perf profiler trampoline; Linux only, and ValueError where the build lacks the backend. Python 3.12+ |
sys.is_stack_trampoline_active() |
O(1) | O(1) | Reads one flag; Python 3.12+ |
sys.monitoring¶
Python 3.12+. Six tool slots, each with its own callbacks and event mask.
| Operation | Time | Space | Notes |
|---|---|---|---|
sys.monitoring.use_tool_id(tool_id, name) |
O(1) | O(1) | Claims one of the ids 0 to 5; ValueError if it is already in use |
sys.monitoring.get_tool(tool_id) |
O(1) | O(1) | The registered name, or None |
sys.monitoring.register_callback(tool_id, event, func) |
O(1) | O(1) | Swaps one slot and hands back the previous callback |
sys.monitoring.set_events(tool_id, event_set) |
O(f + C) | O(C) | Pauses every thread, walks all their frames, and instruments the distinct code objects it finds; anything not currently executing is instrumented when it is next entered |
sys.monitoring.get_events(tool_id) |
O(1) | O(1) | One bitmask read |
sys.monitoring.set_local_events(tool_id, code, event_set) |
O(c) | O(c) | Instruments that one code object, c = its bytecode |
sys.monitoring.get_local_events(tool_id, code) |
O(1) | O(1) | One bitmask read |
sys.monitoring.restart_events() |
O(f + C) | O(C) | The same walk as set_events, so locations a callback turned off start firing again |
sys.monitoring.clear_tool_id(tool_id) |
O(f + C) | O(1) | Drops the tool's callbacks and clears its events, keeping the slot; Python 3.14+ |
sys.monitoring.free_tool_id(tool_id) |
O(f + C) | O(1) | From Python 3.14 it clears the events as above and then releases the slot. On 3.12 and 3.13 it releases the name and nothing else, at O(1) - the callbacks stay registered and instrumented code keeps calling them |
sys.monitoring.events, sys.monitoring.DISABLE, sys.monitoring.MISSING |
O(1) | O(1) | The event namespace, the value a callback returns to stop an event firing at that location, and the value handed to a callback for an argument the interpreter does not have |
sys.monitoring.DEBUGGER_ID, sys.monitoring.COVERAGE_ID, sys.monitoring.PROFILER_ID, sys.monitoring.OPTIMIZER_ID |
O(1) | O(1) | Conventional slot numbers |
Objects and memory¶
| Operation | Time | Space | Notes |
|---|---|---|---|
sys.getsizeof(object, default) |
O(1) | O(1) | Calls object.__sizeof__(), adding the GC header where the type has one; O(1) for builtin types, and whatever a custom __sizeof__ costs otherwise. It never looks inside a container |
sys.getrefcount(object) |
O(1) | O(1) | Reads the refcount field |
sys.intern(string) |
O(len(s)) | O(1) | Hashes and looks the string up, and O(1) when it is already interned. The table holds no reference of its own, so keep the result alive or the entry goes with it - except on Python 3.12.0 to 3.12.6, where the string is immortal instead |
sys._is_interned(string) |
O(1) | O(1) | Reads a flag on the object instead of searching the table, so it does not pay the row above; Python 3.13+ |
sys._is_immortal(object) |
O(1) | O(1) | One refcount comparison; Python 3.14+ |
sys.getunicodeinternedsize() |
O(1) | O(1) | A stored count; Python 3.12+ |
sys.getallocatedblocks() |
O(i) | O(1) | Sums one counter per live interpreter, so O(1) in an ordinary single-interpreter program; from Python 3.13 it pauses every thread while it reads them |
sys._clear_type_cache() |
O(1) | O(1) | Walks a method cache of fixed size; deprecated from Python 3.13 |
sys._clear_internal_caches() |
O(x) | O(1) | The same clear, plus invalidating the x live tier-2 executors; Python 3.13+ |
sys._debugmallocstats() |
O(a) | O(1) | Writes a per-size-class summary to stderr; a = the allocator's arenas |
sys.get_int_max_str_digits(), sys.set_int_max_str_digits(maxdigits) |
O(1) | O(1) | One field; what it caps is the quadratic work in int(str) and str(int). Python 3.10.7+ |
sys.getobjects(limit, type) |
O(objects) | O(limit) | Walks every tracked object; only exists in a Py_TRACE_REFS debug build |
sys.maxsize, sys.maxunicode |
O(1) | O(1) | Build-time constants |
Modules and the import path¶
| Operation | Time | Space | Notes |
|---|---|---|---|
sys.modules |
O(1) | O(1) | The import system's own dict, not a copy of it |
sys.modules[name], name in sys.modules |
O(1) avg | O(1) | An ordinary dict, so O(m) in the worst case if every key collides |
Iterating sys.modules |
O(m) | O(1) | |
sys.path |
O(1) | O(1) | The list the path finder reads on every import miss |
sys.path.append(entry) |
O(1) amortized | O(1) | O(e) on the resize |
sys.path.insert(0, entry) |
O(e) | O(1) | Shifts every existing entry |
entry in sys.path |
O(e) | O(1) | Linear scan |
sys.meta_path |
O(1) | O(1) | List read; an import that misses costs one find_spec() per finder on it |
sys.path_hooks |
O(1) | O(1) | List read; tried in order for a sys.path entry that is not in the cache below |
sys.path_importer_cache |
O(1) avg | O(1) | Dict keyed by path entry, so the sys.path_hooks scan is paid once per entry |
sys.builtin_module_names |
O(1) | O(1) | A tuple, so name in sys.builtin_module_names is O(b) |
sys.stdlib_module_names |
O(1) | O(1) | A frozenset, so membership is O(1) average |
sys.dont_write_bytecode, sys.pycache_prefix |
O(1) | O(1) | Where compiled bytecode goes, and whether it is written at all |
Standard streams and interactive hooks¶
| Operation | Time | Space | Notes |
|---|---|---|---|
sys.stdin, sys.stdout, sys.stderr |
O(1) | O(1) | Attribute reads, and rebinding one is an O(1) assignment; writing costs the text written |
sys.__stdin__, sys.__stdout__, sys.__stderr__ |
O(1) | O(1) | The originals, kept so a replacement can be undone |
sys.displayhook(value), sys.__displayhook__ |
O(r) plus __repr__ |
O(r) | Writes repr(value) to sys.stdout and binds builtins._ to the value; r is what __repr__ returned, and producing it costs whatever the object charges. None returns immediately, writing nothing and leaving builtins._ alone |
sys.breakpointhook(*args, **kwargs), sys.__breakpointhook__ |
O(1) plus the hook | O(1) | What breakpoint() calls; the default imports pdb on first use and then waits for input, and PYTHONBREAKPOINT names what runs instead |
sys.ps1, sys.ps2 |
O(1) | O(1) | Bound only in interactive mode |
sys.__interactivehook__ |
O(1) plus the hook | O(1) | Called once at interactive startup; site installs the one that loads readline history |
Asynchronous generator and coroutine hooks¶
| Operation | Time | Space | Notes |
|---|---|---|---|
sys.get_asyncgen_hooks() |
O(1) | O(1) | Builds a two-field named tuple from the thread state |
sys.set_asyncgen_hooks(firstiter, finalizer) |
O(1) | O(1) | Per thread; firstiter runs once per asynchronous generator, not once per step of it |
sys.get_coroutine_origin_tracking_depth() |
O(1) | O(1) | One field |
sys.set_coroutine_origin_tracking_depth(depth) |
O(1) | O(1) | Constant to set, but while it is non-zero every coroutine created walks up to depth frames |
Encodings and dynamic loading¶
| Operation | Time | Space | Notes |
|---|---|---|---|
sys.getdefaultencoding(), sys.getfilesystemencoding(), sys.getfilesystemencodeerrors() |
O(1) | O(1) | Strings decided at startup and handed back |
sys.getdlopenflags(), sys.setdlopenflags(n) |
O(1) | O(1) | Unix only; the flags extension loading passes to dlopen() |
Interpreter configuration¶
| Operation | Time | Space | Notes |
|---|---|---|---|
sys.version_info, sys.version_info.major, sys.version_info.minor, sys.version_info.micro, sys.version_info.releaselevel, sys.version_info.serial |
O(1) | O(1) | A struct sequence; each field is an index read |
sys.implementation, sys.implementation.name, sys.implementation.version, sys.implementation.hexversion, sys.implementation.cache_tag, sys.implementation.supports_isolated_interpreters |
O(1) | O(1) | Fixed at startup; supports_isolated_interpreters is Python 3.14+ |
sys.flags |
O(1) | O(1) | The command-line and environment switches, fixed at startup - except int_max_str_digits, which set_int_max_str_digits() rewrites from Python 3.14 |
sys.flags.debug, sys.flags.inspect, sys.flags.interactive, sys.flags.optimize, sys.flags.dont_write_bytecode, sys.flags.no_user_site, sys.flags.no_site, sys.flags.ignore_environment, sys.flags.verbose, sys.flags.bytes_warning, sys.flags.quiet, sys.flags.hash_randomization, sys.flags.isolated, sys.flags.dev_mode, sys.flags.utf8_mode, sys.flags.warn_default_encoding, sys.flags.safe_path, sys.flags.int_max_str_digits, sys.flags.gil, sys.flags.thread_inherit_context, sys.flags.context_aware_warnings |
O(1) | O(1) | Index reads; safe_path is Python 3.11+, gil 3.13+, and thread_inherit_context and context_aware_warnings 3.14+ |
sys.float_info, sys.float_info.max, sys.float_info.max_exp, sys.float_info.max_10_exp, sys.float_info.min, sys.float_info.min_exp, sys.float_info.min_10_exp, sys.float_info.dig, sys.float_info.mant_dig, sys.float_info.epsilon, sys.float_info.radix, sys.float_info.rounds |
O(1) | O(1) | The C float.h limits, as index reads |
sys.int_info, sys.int_info.bits_per_digit, sys.int_info.sizeof_digit, sys.int_info.default_max_str_digits, sys.int_info.str_digits_check_threshold |
O(1) | O(1) | How int stores digits, and the limit described under set_int_max_str_digits |
sys.hash_info, sys.hash_info.width, sys.hash_info.modulus, sys.hash_info.inf, sys.hash_info.nan, sys.hash_info.imag, sys.hash_info.algorithm, sys.hash_info.hash_bits, sys.hash_info.seed_bits, sys.hash_info.cutoff |
O(1) | O(1) | The parameters of the hash the dict and set bounds assume is O(1) |
sys.thread_info, sys.thread_info.name, sys.thread_info.lock, sys.thread_info.version |
O(1) | O(1) | The threading implementation, fixed at build time |
Runtime data attributes¶
| Operation | Time | Space | Notes |
|---|---|---|---|
sys.platform, sys.version, sys.hexversion, sys.api_version, sys.byteorder, sys.copyright, sys.float_repr_style, sys.abiflags |
O(1) | O(1) | Plain attribute reads, computed once at startup |
sys.executable, sys.prefix, sys.exec_prefix, sys.base_prefix, sys.base_exec_prefix, sys.platlibdir |
O(1) | O(1) | Paths resolved during startup |
sys.argv, sys.orig_argv |
O(1) | O(1) | Lists built once; orig_argv keeps the interpreter's own options, which argv drops. Python 3.10+ |
sys.warnoptions, sys._xoptions |
O(1) | O(1) | The -W and -X options as given on the command line |
sys.is_finalizing(), sys._is_gil_enabled() |
O(1) | O(1) | One flag each; _is_gil_enabled is Python 3.13+ |
sys._jit.is_available(), sys._jit.is_enabled(), sys._jit.is_active() |
O(1) | O(1) | Flag reads for the tier-2 JIT; Python 3.14+ |
Platform-specific and remote debugging¶
| Operation | Time | Space | Notes |
|---|---|---|---|
sys.getwindowsversion() |
O(1) | O(1) | Windows only |
sys.dllhandle, sys.winver |
O(1) | O(1) | Windows only |
sys._enablelegacywindowsfsencoding() |
O(1) | O(1) | Windows only; deprecated from Python 3.13 |
sys.getandroidapilevel() |
O(1) | O(1) | Android only |
sys._emscripten_info, sys._emscripten_info.emscripten_version, sys._emscripten_info.runtime, sys._emscripten_info.pthreads, sys._emscripten_info.shared_memory |
O(1) | O(1) | Emscripten only |
sys.remote_exec(pid, script) |
O(1) | O(1) | Checks the script is readable, then writes its path into the target process; the target compiles and runs it at its next safe point, so the script's own cost lands there. Python 3.14+ |
sys.is_remote_debug_enabled() |
O(1) | O(1) | Whether this interpreter will accept the row above; Python 3.14+ |
Walking the Stack¶
sys._getframe(depth) follows one link per level, so asking for the caller is cheap and asking
for the frame 500 levels up is not. sys._current_frames() and sys._current_exceptions() cost
one entry per live thread instead, whatever each stack looks like.
import sys
def innermost():
# O(d) - two links up from here
return sys._getframe(2).f_code.co_name
def middle():
return innermost()
def outermost():
return middle()
assert outermost() == 'outermost'
frames = sys._current_frames() # O(t) - one entry per live thread
exceptions = sys._current_exceptions() # O(t) - the same threads
assert set(exceptions) == set(frames)
try:
sys._getframe(1_000_000)
except ValueError as error:
assert 'call stack is not deep enough' in str(error)
else:
raise AssertionError('the walk ran off the bottom without complaining')
The Recursion Limit¶
Reading the limit is one field. Setting it has been a write through every thread state since Python 3.11, which is what the O(t) in the table is.
import sys
original = sys.getrecursionlimit() # O(1)
sys.setrecursionlimit(original + 500) # O(t) - one pass over the thread states
assert sys.getrecursionlimit() == original + 500
try:
sys.setrecursionlimit(1)
except RecursionError as error:
assert 'limit is too low' in str(error)
else:
raise AssertionError('a limit below the current depth was accepted')
finally:
sys.setrecursionlimit(original)
Interning Strings¶
The first sys.intern() of a string hashes it and compares it against the table entry it lands
on, so it costs the string's length. Handing back a string that is already interned costs
nothing of the sort, and sys._is_interned() answers the question without the lookup at all.
import sys
first = sys.intern(''.join(['inter', 'ning'])) # O(len(s)) - hashed and looked up
second = sys.intern(''.join(['inter', 'ning'])) # O(len(s)), then the table hit
assert first is second
assert sys.intern(first) is first # O(1) - already interned
if sys.version_info >= (3, 13):
assert sys._is_interned(first) is True # O(1) - a flag read
assert sys._is_interned(''.join(['not', '-interned'])) is False
Interning is permanent on Python 3.12.0 to 3.12.6
Across those seven patch releases an interned string is immortal: dropping every reference to it does not give the memory back, so interning many distinct strings grows the process for good. 3.12.7 restored the mortal table every other supported release has.
Tracing, Profiling and Monitoring¶
Installing a Tracer, and Then Paying for It¶
sys.settrace() and sys.setprofile() store one function, and from Python 3.12 they also go
through the sys.monitoring machinery below, so installing one pauses every thread and
instruments the code objects on their stacks. Either way the bill that dominates comes
afterwards, once per traced event - for settrace(), once per line executed.
import sys
def work(count):
total = 0
for value in range(count):
total += value
return total
lines = []
def tracer(frame, event, arg):
if event == 'line':
lines.append(frame.f_lineno)
return tracer
original = sys.gettrace() # O(1)
sys.settrace(tracer) # O(f + C) from 3.12, O(1) before it
try:
work(5) # one tracer call per line executed
finally:
sys.settrace(original) # O(f + C) from 3.12, O(1) before it
assert len(lines) > 5, lines
assert sys.gettrace() is original
Instrumentation Follows the Code Object¶
A sys.monitoring callback still runs once per event, but the event is reached by instrumented
bytecode rather than by a per-frame trace function, so a tool can ask for one event in one code
object and pay nothing anywhere else. What switching it on costs is the code it rewrites:
set_local_events() instruments the one code object it is handed, and set_events() pauses the
interpreter and instruments everything currently on a stack, leaving the rest to be instrumented
as it is entered.
import sys
if sys.version_info >= (3, 12):
monitoring = sys.monitoring
tool = monitoring.DEBUGGER_ID
monitoring.use_tool_id(tool, 'page-example') # O(1)
seen = []
monitoring.register_callback( # O(1)
tool, monitoring.events.LINE, lambda code, lineno: seen.append(lineno)
)
def counted():
total = 0
for value in range(3):
total += value
return total
# O(c) - instruments this one code object, c = its bytecode
monitoring.set_local_events(tool, counted.__code__, monitoring.events.LINE)
assert counted() == 3
monitoring.set_local_events(tool, counted.__code__, 0) # O(c)
assert monitoring.get_tool(tool) == 'page-example' # O(1)
monitoring.free_tool_id(tool) # O(f + C) from 3.14
assert len(seen) > 3, seen
Auditing¶
A sys.audit() call runs the installed hooks in order until one of them raises, so the cost of
an event is the number of hooks times whatever each one does. Adding a hook costs the same walk,
because addaudithook() raises its own event first, and there is no way to undo it.
import sys
events = []
sys.addaudithook(lambda event, args: events.append(event)) # O(h), and permanent
assert events == [] # the hook is installed after its own event is raised
sys.addaudithook(lambda event, args: None) # O(h) - the hook above sees this one arrive
assert events == ['sys.addaudithook']
sys.audit('page.probe', 1, 2) # O(h) - each hook in turn, until one raises
assert events[-1] == 'page.probe'
assert not hasattr(sys, 'removeaudithook')
Modules and the Import Path¶
sys.modules is the import system's cache, and looking a name up in it is the dict lookup that
makes a repeated import cheap. The two module-name collections beside it are not
interchangeable: one is a tuple and one is a frozenset.
import json
import sys
assert sys.modules['json'] is json # O(1) average - the import cache itself
assert type(sys.modules) is dict
assert 'sys' in sys.builtin_module_names # O(b) - a tuple, scanned linearly
assert 'json' in sys.stdlib_module_names # O(1) average - a frozenset
sys.path is a list, so where you add an entry decides what it costs. Prepending k entries one
at a time shifts everything in front of each new one - O(ke + k*2), since the list is longer
every time round. A slice assignment shifts each entry once, at O(e + k), and keeps the list
object the import system already holds.
import sys
original = list(sys.path)
additions = ['/one', '/two', '/three']
for entry in additions:
sys.path.insert(0, entry) # the list grows as it goes, so the loop is O(k*e + k**2)
assert sys.path[:3] == ['/three', '/two', '/one']
sys.path[:] = original
identity = id(sys.path)
sys.path[:0] = additions # O(e + k) once, and the same list object
assert id(sys.path) == identity
assert sys.path[:3] == additions
sys.path[:] = original
assert '/one' not in sys.path # O(e) - a linear scan
Replacing the Standard Streams¶
sys.stdout is a name like any other, so redirecting it is an assignment and restoring it is
another. What costs anything is the text that goes through the replacement.
import builtins
import io
import sys
original = sys.stdout
captured = io.StringIO()
sys.stdout = captured # O(1)
print('redirected')
sys.stdout = original # O(1)
assert captured.getvalue() == 'redirected\n'
shown = io.StringIO()
sys.stdout = shown
sys.displayhook([1, 2, 3]) # O(r) - r = characters in repr(value)
sys.stdout = original
assert shown.getvalue() == '[1, 2, 3]\n'
assert builtins._ == [1, 2, 3]
Sizing an Object¶
sys.getsizeof() asks the object and nobody else. A list of one large object is the same size
as a list of one small one, because what the list holds is a pointer.
import sys
assert sys.getsizeof([b'x']) == sys.getsizeof([b'x' * 1_000_000]) # O(1)
class Measured:
def __sizeof__(self):
return 42
assert sys.getsizeof(Measured()) >= 42 # the GC header is added to what it returned
class Opaque:
__slots__ = ()
__sizeof__ = None
try:
sys.getsizeof(Opaque())
except TypeError as error:
assert 'not callable' in str(error)
else:
raise AssertionError('an object with no __sizeof__ was measured anyway')
assert sys.getsizeof(Opaque(), 99) == 99 # the default is returned instead
Asynchronous Generator Hooks¶
sys.set_asyncgen_hooks() stores two functions for the current thread. firstiter is called
once for each asynchronous generator, when it is first stepped - not once per step - so its
cost scales with generators created, not with iterations.
import sys
started = []
sys.set_asyncgen_hooks(firstiter=started.append) # O(1), per thread
async def numbers():
yield 1
yield 2
def step(awaitable):
try:
awaitable.send(None)
except StopIteration as stop:
return stop.value
raise AssertionError('the step did not finish')
generator = numbers()
assert step(generator.__anext__()) == 1
assert step(generator.__anext__()) == 2
assert len(started) == 1 # once per generator, not once per step
sys.set_asyncgen_hooks(None, None) # O(1)
assert sys.get_asyncgen_hooks() == (None, None)
Common Patterns¶
Reporting the Caller's Location¶
import sys
def caller_location(depth=1):
frame = sys._getframe(depth + 1) # O(d) - one link per level
return frame.f_code.co_name, frame.f_lineno
def helper():
return caller_location()
def entry():
return helper()
name, _ = entry()
assert name == 'entry'
Printing a Traceback Without Raising¶
import io
import sys
def inner():
raise ValueError('boom')
def outer():
inner()
report = io.StringIO()
original = sys.stderr
sys.stderr = report
try:
outer()
except ValueError:
sys.excepthook(*sys.exc_info()) # O(d + s + m) - three distinct frames, so three entries
finally:
sys.stderr = original
text = report.getvalue()
assert text.count(' File "') == 3 # this block, outer, inner
assert text.rstrip().endswith('ValueError: boom')
Performance Best Practices¶
✅ Do:
- Reach for
sys._getframe(0)orsys._getframe(1)and let something else walk further; the cost is the depth you ask for - Prepend to
sys.pathwith a slice assignment: O(e + k) once, rather than O(ke + k*2) for the same entries inserted one at a time - Use
sys.stdlib_module_namesfor a membership test andsys.builtin_module_namesonly when you need the tuple itself - Prefer
sys.exception()tosys.exc_info()on Python 3.11+ when you only want the exception - Scope
sys.monitoringwithset_local_events()where one code object is what you are after
❌ Avoid:
- Leaving a
settrace()function installed: on the thread that installed it, it is called for every frame entered and, unless it declines that frame by returningNone, for every line and every exception inside it - Leaving a
setprofile()function installed: fewer events - calls and returns, Python and C - but the return value is ignored, so a frame cannot be declined. From Python 3.12 installing or removing either one stops every thread as well - Installing an audit hook you may want back out; nothing removes one
- Interning strings you do not keep a reference to, and on Python 3.12.0 to 3.12.6 interning many distinct strings at all
- Treating
sys.getsizeof()as the size of a container's contents - it is the size of the container - Calling
sys.getallocatedblocks()in a hot path; from Python 3.13 it pauses every thread to read its counters
Version Notes¶
- Python 3.10.7+:
sys.set_int_max_str_digits()andsys.get_int_max_str_digits()added, withsys.flags.int_max_str_digitsandsys.int_info.default_max_str_digits - Python 3.11+:
sys.exception()andsys.flags.safe_pathadded, andsys.setrecursionlimit()began writing through to every thread state, which is where its O(t) comes from - Python 3.12+:
sys.monitoring,sys.getunicodeinternedsize(),sys._getframemodulename(),sys.last_excand the stack trampoline calls added;sys._current_exceptions()maps each thread to an exception rather than to anexc_info()triple;sys.settrace()andsys.setprofile()moved onto the monitoring machinery, so installing one went from O(1) to O(f + C) - Python 3.12.0 to 3.12.6:
sys.intern()makes the string immortal, so interning many distinct strings never gives the memory back; 3.12.7 restored the mortal table - Python 3.13+:
sys._is_interned(),sys._is_gil_enabled(),sys.flags.gilandsys._clear_internal_caches()added, andsys._clear_type_cache()documented as deprecated in favour of the last of those - Python 3.14+: calling
sys._clear_type_cache()raisesDeprecationWarning;sys.remote_exec(),sys.is_remote_debug_enabled(),sys._is_immortal(),sys._jitandsys.monitoring.clear_tool_id()added, andsys.monitoring.free_tool_id()began clearing the tool's events and callbacks instead of only releasing the name - All Python 3: an audit hook cannot be removed once added, so install one for the life of the process or not at all
Related Modules¶
- traceback - formats the frames
sys.excepthookprints - threading - the threads the O(t) rows count
- importlib - the finders behind
sys.meta_pathandsys.path_hooks - gc - reaches the objects
sys.getsizeof()deliberately does not - tracemalloc - allocation totals, where
sys.getallocatedblocks()gives one number