Formatting run info no longer includes gevent.local.local
objects that have no value in the greenlet. See issue #1275.
Fixed negative length in pywsgi’s Input read functions for non chunked body. Reported in issue #1274 by tzickel.
Upgrade libuv from 1.22.0 to 1.23.2.
Fix opening files in text mode in CPython 2 on Windows by patching libuv. See issue #1282 reported by wiggin15.
gevent now depends on greenlet 0.4.14 or above. gevent binary wheels for 1.3.5 and below must have greenlet 0.4.13 installed on Python 3.7 or they will crash. Reported by Alexey Stepanov in issue #1260 and pkittenis in issue #1261.
gevent.local.local
subclasses correctly supports
@staticmethod
functions. Reported by Brendan Powers in
issue #1266.
Update the bundled libuv from 1.20.1 to 1.22.0.
Test Python 3.7 on Appveyor. Fix the handling of Popen’s
close_fds
argument on 3.7.
Update Python versions tested on Travis, including PyPy to 6.0. See issue #1195.
gevent.queue
imports _PySimpleQueue
instead of
SimpleQueue
so that it doesn’t block the event loop.
gevent.monkey.patch_all()
makes this same substitution in
queue
. This fixes issues with
concurrent.futures.ThreadPoolExecutor
as well. Reported in
issue #1248 by wwqgtxx and issue #1251 by pyld.
gevent.socket.socket.connect()
doesn’t pass the port (service)
to socket.getaddrinfo()
when it resolves an AF_INET
or
AF_INET6
address. (The standard library doesn’t either.) This
fixes an issue on Solaris. Reported in issue #1252 by wiggin15.
gevent.socket.socket.connect()
works with more address
families, notably AF_TIPC, AF_NETLINK, AF_BLUETOOTH, AF_ALG and AF_VSOCK.
Be more careful about issuing MonkeyPatchWarning
for ssl
imports. Now, we only issue it if we detect the one specific
condition that is known to lead to RecursionError. This may produce
false negatives, but should reduce or eliminate false positives.
Based on measurements and discussion in issue #1233, adjust the
way gevent.pywsgi
generates HTTP chunks. This is intended to
reduce network overhead, especially for smaller chunk sizes.
Additional slight performance improvements in gevent.pywsgi
.
See pull request #1241.
gevent.sleep()
updates the loop’s notion of the current time
before sleeping so that sleep duration corresponds more closely to
elapsed (wall clock) time. gevent.Timeout
does the same.
Reported by champax and FoP in issue #1227.
Fix an UnboundLocalError
in SSL servers when wrapping a socket
throws an error. Reported in issue #1236 by kochelmonster.
Fix a packaging error in manylinux binary wheels that prevented some imports from working. See issue #1219.
Allow weak refeneces to gevent.queue.Queue
. Reported in
issue #1217 by githrdw.
Allow weak references to gevent.event.Event
. Reported in
issue #1211 by Matias Guijarro.
Fix embedded uses of gevent.Greenlet.spawn()
, especially under
uwsgi. Reported in issue #1212 by Kunal Gangakhedkar.
Fix gevent.os.nb_write()
and gevent.os.nb_read()
not
always closing the IO event they opened in the event of an
exception. This would be a problem especially for libuv.
Python 3.7 passes the automated memory leak checks. See issue #1197.
Update autoconf’s config.guess and config.sub to the latest versions for c-ares and libev.
gevent.local.local
subclasses that mix-in ABCs can be instantiated.
Reported in issue #1201 by Bob Jordan.
On Windows, CFFI is now a dependency so that the libuv backend really can be used by default.
Fix a bug detecting whether we can use the memory monitoring features when psutil is not installed.
gevent.subprocess.Popen
uses /proc/self/fd
(on Linux) or
/dev/fd
(on BSD, including macOS) to find the file descriptors
to close when close_fds
is true. This matches an optimization
added to Python 3 (and backports it to Python 2.7), making process
spawning up to 9 times faster. Also, on Python 3, since Python 3.3
is no longer supported, we can also optimize the case where
close_fds
is false (not the default), making process spawning up
to 38 times faster. Initially reported in issue #1172 by Ofer Koren.
The bundled libuv is now 1.20.1, up from 1.19.2. See issue #1177.
The long-deprecated and undocumented module gevent.wsgi
was removed.
Add gevent.util.assert_switches
to build on the monitoring
functions. Fixes issue #1182.
A started monitor thread for the active hub now survives a fork. See issue #1185.
The greenlet tracer functions used for the various monitoring capabilities are now compiled with Cython for substantially lower overhead. See pull request #1190.
libuv now collects all pending watchers and runs their callbacks at the end of the loop iteration using UV_RUN_ONCE. This eliminates the need to patch libuv to be greenlet-safe. It also means that zero-duration timer watchers are actual timer watchers again (instead of being turned into check watchers); newly added zero-duration timers cannot block the event loop because they won’t be run until a safe time.
Python 3.7.0b4 is now the tested and supported version of Python 3.7. PyPy 6.0 has been tested, although CI continues to use 5.10.
Cython 0.28.2 is now used to build gevent from a source checkout.
The bundled libuv is now 1.19.2, up from 1.18.0.
Travis CI tests on Python 3.7.0b3.
Windows now defaults to the libuv backend if CFFI is installed. See issue #1163.
On Python 2, when monkey-patching threading.Event
, also
monkey-patch the underlying class, threading._Event
. Some code
may be type-checking for that. See issue #1136.
Fix libuv io watchers polling for events that only stopped watchers are interested in, reducing CPU usage. Reported in issue #1144 by wwqgtxx.
Add additional optimizations for spawning greenlets, making it faster than 1.3a2.
Use strongly typed watcher callbacks in the libuv CFFI extensions. This prevents dozens of compiler warnings.
When gevent prints a timestamp as part of an error message, it is now in UTC format as specified by RFC3339.
Threadpool threads that exit now always destroy their hub (if one was created). This prevents some forms of resource leaks (notably visible as blocking functions reported by the new monitoring abilities).
Hub objects now include the value of their name
attribute in
their repr.
Pools for greenlets and threads have lower overhead, especially for
map
. See pull request #1153.
The undocumented, internal implementation classes IMap
and
IMapUnordered
classes are now compiled with Cython, further
reducing the overhead of [Thread]Pool.imap
.
The classes gevent.event.Event
and gevent.event.AsyncResult
are compiled with Cython for improved performance, as is the
gevent.queue
module and gevent.hub.Waiter
and certain
time-sensitive parts of the hub itself. Please report any
compatibility issues.
python -m gevent.monkey <script>
accepts more values for
<script>
, including paths to packages or compiled bytecode.
Reported in issue #1157 by Eddie Linder.
Add a simple event framework for decoupled communication. It uses
zope.event
if that is installed.
gevent.monkey
has support for plugins in the form of event
subscribers and setuptools entry points. See pull request #1158 and
issue #1162. setuptools must be installed at runtime for its entry
points to function.
Introduce the configuration variable
gevent.config.track_greenlet_tree
(aka
GEVENT_TRACK_GREENLET_TREE
) to allow disabling the greenlet tree
features for applications where greenlet spawning is performance
critical. This restores spawning performance to 1.2 levels.
Add an optional monitoring thread for each hub. When enabled, this
thread (by default) looks for greenlets that block the event loop
for more than 0.1s. You can add your own periodic monitoring
functions to this thread. Set GEVENT_MONITOR_THREAD_ENABLE
to
use it, and GEVENT_MAX_BLOCKING_TIME
to configure the blocking
interval.
The monitoring thread emits events when it detects certain conditions, like loop blocked or memory limits exceeded.
Add settings for monitoring memory usage and emitting events when a
threshold is exceeded and then corrected. gevent currently supplies
no policy for what to do when memory exceeds the configured limit.
psutil
must be installed to use this. See pull request #1150.
Cython 0.28b1 or later is now required to build gevent from a source checkout (Cython is not required to build a source distribution from PyPI).
Update c-ares to 1.14.0. See issue #1105.
gevent now requires the patched version of libuv it is distributed with. Building gevent with a non-embedded libuv, while not previously supported, is not possible now. See issue #1126.
Travis CI tests on Python 3.7.0b2 and PyPy 2.7 5.10.0 and PyPy 3.5 5.10.1.
Fix building from a source distribution (PyPI) without Cython installed.
Add the dnspython
resolver as a lightweight alternative to
c-ares. It is generally faster than c-ares and is supported on PyPy.
c-ares may be deprecated in the future. See pull request #1088 and
issue #1103.
Add the module gevent.time
that can be imported instead of
time
, much like gevent.socket
can be imported instead
of socket
. It contains gevent.sleep
. This aids
monkey-patching.
Simple subclasses of gevent.local.local
now have the same
(substantially improved) performance characteristics of plain
gevent.local.local
itself, making them 2 to 3 times faster than
before. See pull request #1117. If there are any compatibility
problems, please open issues.
Greenlet objects now keep track of their spawning parent greenlet and the code location that spawned them, in addition to maintaining a “spawn tree local” mapping. This adds some runtime overhead in relative terms, but absolute numbers are still relatively small. Based on a proposal from PayPal and comments by Mahmoud Hashemi and Kurt Rose. See issue #755 and pull request #1115. As always, feedback is appreciated.
Greenlet objects now have a minimal_ident
property. It functions
similarly to Thread.ident
or id
by uniquely identifying the
greenlet object while it remains alive, and it can be reused after
the greenlet object is dead. It is different in that it is small and
sequential. Based on a proposal from PayPal and comments by Mahmoud
Hashemi and Kurt Rose. See issue #755. As always, feedback is
appreciated.
gevent.Greenlet
objects now have a gevent.Greenlet.name
attribute that is included in the default repr.
Include the values of gevent.local.local
objects associated with
each greenlet in gevent.util.format_run_info
.
Add gevent.util.GreenletTree
to visualize the greenlet tree. This
is used by gevent.util.format_run_info
.
Make gevnt.subprocess.Popen
accept the restore_signals
keyword argument on all versions of Python, and on Python 2 have it
default to false. It previously defaulted to true on all versions;
now it only defaults to true on Python 3. The standard library in
Python 2 does not have this argument and its behaviour with regards
to signals is undocumented, but there is code known to rely on
signals not being restored under Python 2. Initial report and patch
in pull request #1063 by Florian Margaine.
Allow gevent.subprocess.Popen
to accept the keyword
arguments pass_fds
and start_new_session
under Python 2.
They have always had the same default as Python 3, namely an empty
tuple and false, but now are accessible to Python 2.
Support the capture_output
argument added to Python 3.7 in
gevent.subprocess.run()
.
Centralize all gevent configuration in an object at
gevent.config
, allowing for gevent to be configured through code
and not necessarily environment variables, and also provide a
centralized place for documentation. See issue #1090.
The new GEVENT_CORE_CFFI_ONLY
environment variable has been
replaced with the pre-existing GEVENT_LOOP
environment
variable. That variable may take the values libev-cext
,
libev-cffi
, or libuv-cffi
, (or be a list in preference
order, or be a dotted name; it may also be assigned to an
object in Python code at gevent.config.loop
).
The GEVENTARES_SERVERS
environment variable is deprecated in
favor of GEVENT_RESOLVER_SERVERS
. See issue #1103.
Fix calling shutdown
on a closed socket. It was raising
AttributeError
, now it once again raises the correct
socket.error
. Reported in issue #1089 by André Cimander.
Fix an interpreter crash that could happen if two or more loop
objects referenced the default event loop and one of them was
destroyed and then the other one destroyed or (in the libev C
extension implementation only) deallocated (garbage collected). See
issue #1098.
Fix a race condition in libuv child callbacks. See issue #1104.
The internal, undocumented module gevent._threading
has been
simplified.
The internal, undocumented class gevent._socket3._fileobject
has
been removed. See issue #1084.
Simplify handling of the libev default loop and the destroy()
method. The default loop, when destroyed, can again be requested and
it will regenerate itself. The default loop is the only one that can
receive child events.
Make gevent.socket.socket.sendall()
up to ten times faster on
PyPy3, through the same change that was applied in gevent 1.1b3 for PyPy2.
Be more careful about issuing a warning about patching SSL on Python 2. See issue #1108.
Signal handling under PyPy with libuv is more reliable. See issue #1112.
The gevent.greenlet
module is now compiled with Cython to
offset any performance decrease due to issue #755. Please open
issues for any compatibility concerns. See pull request #1115 and pull request #1120.
On CPython, allow the pure-Python implementations of
gevent.Greenlet
, gevent.local
and gevent.lock
to be
used when the environment variable PURE_PYTHON
is set. This is
not recommended except for debugging and testing. See issue #1118.
gevent.select.poll.poll()
now interprets a timeout of -1 the
same as a timeout of None as the standard requires. Previously,
on libuv this was interpreted the same as a timeout of 0. In
addition, all timeout values less than zero are interpreted like
None (as they always were under libev). See issue #1127.
Monkey-patching now defaults to patching threading.Event
.
gevent is now built and tested with Cython 0.27. This is required for Python 3.7 support.
Update c-ares to 1.13.0. See issue #990.
Add initial support for Python 3.7a3. It has the same level of support as Python 3.6.
Using unreleased Cython 0.28 and greenlet 0.4.13; requires Python 3.7a3.
The async
functions and classes have been renamed to
async_
due to async
becoming a keyword in Python 3.7.
Aliases are still in place for older versions. See issue #1047.
gevent is now tested on Python 3.6.4. This includes the following fixes and changes:
Errors raised from gevent.subprocess
will have a
filename
attribute set.
The threading.Timer
class is now monkey-patched and can
be joined. Previously on Python 3.4 and above, joining a Timer
would hang the process.
gevent.ssl.SSLSocket.unwrap()
behaves more like the standard
library, including returning a SSLSocket and allowing certain
timeout-related SSL errors to propagate. The added standard
library tests test_ftplib.py
now passes.
gevent.subprocess.Popen
accepts a “path-like object” for
the cwd parameter on all platforms. Previously this only worked
on POSIX platforms under Python 3.6. Now it also works on Windows under
Python 3.6 (as expected) and is backported to all previous versions.
Linux CI now tests on PyPy3 3.5-5.9.0, updated from PyPy3 3.5-5.7.1. See issue #1001. PyPy2 has been updated to 5.9.0 from 5.7.1, Python 2.7 has been updated to 2.7.14 from 2.7.13, Python 3.4 is updated to 3.4.7 from 3.4.5, Python 3.5 is now 3.5.4 from 3.5.3, and Python 3.6 is now 3.6.4 from 3.6.0.
Drop support for Python 3.3. The documentation has only claimed support for 3.4+ since gevent 1.2 was released, and only 3.4+ has been tested. This merely removes the supporting Trove classifier and remaining test code. See issue #997.
PyPy is now known to run on Windows using the libuv backend, with caveats. See the section on libuv for more information.
Due to security concerns, official support for Python 2.7.8 and earlier (without a modern SSL implementation) has been dropped. These versions are no longer tested with gevent, but gevent can still be installed on them. Supporting code will be removed in the next major version of gevent. See issue #1073.
When building gevent from a source checkout (not a distributed
source distribution), make
is no longer required and the
Makefile
is not used. Neither is an external cython
command.
Instead, the cythonize
function is used, as recommended by
Cython. (The external commands were never required by source
distributions.) See issue #1076.
gevent.local.local
is compiled with Cython on CPython.
The Cython ares ‘channel’ class is no longer declared to be publicly accessible from a named C structure. Doing so caused a conflict with the c-ares header files.
If a single greenlet created and destroyed many
gevent.local.local
objects without ever exiting, there
would be a leak of the function objects intended to clean up the
locals after the greenlet exited. Introduce a weak reference to
avoid that. Reported in issue #981 by Heungsub Lee.
pywsgi also catches and ignores by default
errno.WSAECONNABORTED
on Windows. Initial patch in
pull request #999 by Jan van Valburg.
gevent.subprocess.Popen.communicate()
returns the correct type
of str (not bytes) in universal newline mode under Python 3, or when
an encoding has been specified. Initial patch in pull request #939 by
William Grzybowski.
gevent.subprocess.Popen.communicate()
(and in general,
accessing Popen.stdout
and Popen.stderr
) returns the correct
type of str (bytes) in universal newline mode under Python 2.
Previously it always returned unicode strings. Reported in
issue #1039 by Michal Petrucha.
gevent.fileobject.FileObjectPosix
returns native strings in
universal newline mode on Python 2. This is consistent with what
FileObjectThread
does. See issue #1039.
socket.send()
now catches EPROTOTYPE
on macOS to handle a race
condition during shutdown. Fixed in pull request #1035 by Jay Oster.
gevent.socket.create_connection()
now properly cleans up open
sockets if connecting or binding raises a BaseException
like
KeyboardInterrupt
, greenlet.GreenletExit
or
gevent.timeout.Timeout
. Reported in issue #1044 by
kochelmonster.
Pool.add
now accepts blocking
and timeout
parameters,
which function similarly to their counterparts in Semaphore
.
See pull request #1032 by Ron Rothman.
Defer adjusting the stdlib’s list of active threads until
threading
is monkey patched. Previously this was done when
gevent.threading
was imported. That module is documented to
be used as a helper for monkey patching, so this should generally
function the same, but some applications ignore the documentation
and directly import that module anyway.
A positive consequence is that import gevent.threading, threading;
threading.current_thread()
will no longer return a DummyThread
before monkey-patching. Another positive consequence is that PyPy
will no longer print a KeyError
on exit if
gevent.threading
was imported without monkey-patching.
See issue #984.
Specify the Requires-Python metadata for improved installation support in certain tools (setuptools v24.2.1 or newer is required). See issue #995.
Monkey-patching after the ssl
module has been imported now
prints a warning because this can produce RecursionError
.
gevent.local.local
objects are now approximately 3.5 times faster
getting, setting and deleting attributes on PyPy. This involved
implementing more of the attribute protocols directly. Please open
an issue if you have any compatibility problems. See issue #1020.
gevent.local.local
is compiled with Cython on CPython. It
was already 5 to 6 times faster due to the work on issue #1020,
and compiling it with Cython makes it another 5 to 6 times faster,
for a total speed up of about 35 times. It is now in the same
ballpark as the native threading.local
class. It also uses
one pointer less memory per object, and one pointer less memory per
greenlet. See pull request #1024.
More safely terminate subprocesses on Windows with
gevent.subprocess.Popen.terminate()
. Reported in issue #1023
by Giacomo Debidda.
gevent now uses cffi’s “extern ‘Python’” callbacks. These should be faster and more stable. This requires at least cffi 1.4.0. See issue #1049.
gevent now approximately tries to stick to a scheduling interval
when running callbacks, instead of simply running a count of
callbacks. The interval is determined by
gevent.getswitchinterval()
. On Python 3, this is the same as
the thread switch interval. On Python 2, this defaults to 0.005s and
can be changed with gevent.setswitchinterval()
. This should
result in more fair “scheduling” of greenlets, especially when
gevent.sleep(0)
or other busy callbacks are in use. The interval
is checked every 50 callbacks to keep overhead low. See
issue #1072. With thanks to Arcadiy Ivanov and Antonio Cuni.
Add initial experimental support for using libuv as a backend
instead of libev, controlled by setting the environment variable
GEVENT_CORE_CFFI_ONLY=libuv
before importing gevent. This
suffers a number of limitations compared to libev, notably:
libuv support is not available in the manylinux wheels uploaded to
PyPI. The manylinux specification requires glibc 2.5, while libuv
requires glibc 2.12. Install from source to access libuv on Linux
(e.g., pip’s --no-binary
option).
Timers (such as gevent.sleep
and gevent.Timeout
) only
support a resolution of 1ms (in practice, it’s closer to 1.5ms).
Attempting to use something smaller will automatically increase it
to 1ms and issue a warning. Because libuv only supports
millisecond resolution by rounding a higher-precision clock to an
integer number of milliseconds, timers apparently suffer from more
jitter.
Using negative timeouts may behave differently from libev.
libuv blocks delivery of all signals, so signals are handled using
an (arbitrary) 0.3 second timer. This means that signal handling
will be delayed by up to that amount, and that the longest the
event loop can sleep in the operating system’s poll
call is
that amount. Note that this is what gevent does for libev on
Windows too.
libuv only supports one io watcher per file descriptor, whereas libev and gevent have always supported many watchers using different settings. The libev behaviour is emulated at the Python level, but that adds overhead.
Looping multiple times and expecting events for the same file
descriptor to be raised each time without any data being read or
written (as works with libev) does not appear to work correctly on
Linux when using gevent.select.poll
or a monkey-patched
selectors.PollSelector
.
The build system does not support using a system libuv; the embedded copy must be used. Using setuptools to compile libuv was the most portable method found.
If anything unexpected happens, libuv likes to abort()
the
entire process instead of reporting an error. For example, closing
a file descriptor it is using in a watcher may cause the entire
process to be exited.
There may be occasional otherwise unexplained and hard to duplicate crashes. If you can duplicate a crash, please submit an issue.
This is the only backend that PyPy can use on Windows. As of this
alpha, there are many known issues with non-blocking sockets
(e.g., as used by asyncore
; see test_ftplib.py
) and
sometimes sockets not getting closed in a timely fashion
(apparently; see test_httpservers.py
) and communicating with
subprocesses (it always hangs). Help tracking those down would be
appreciated. Only PyPy2 is tested.
Feedback and pull requests are welcome, especially to address the issues mentioned above.
Other differences include:
The order in which timers and other callbacks are invoked may be different than in libev. In particular, timers and IO callbacks happen in a different order, and timers may easily be off by up to half of the supposed 1ms resolution. See issue #1057.
Starting a timer
watcher does not update the loop’s time by
default. This is because, unlike libev, a timer callback can cause
other timer callbacks to be run if they expire because the loop’s
time updated, without cycling the event loop. See issue #1057.
libev has also been changed to follow this behaviour.
Also see issue #1072.
Timers of zero duration do not necessarily cause the event loop to cycle, as they do in libev. Instead, they may be called immediately. If zero duration timers are added from other zero duration timer callbacks, this can lead the loop to appear to hang, as no IO will actually be done.
To mitigate this issue, loop.timer()
detects attempts to use
zero duration timers and turns them into a check watcher. check
watchers do not support the again
method.
All watchers (e.g., loop.io
) and the Timeout
class have a
close
method that should be called when code is done using the
object (they also function as context managers and a with
statement will automatically close them). gevent does this
internally for sockets, file objects and internal timeouts.
Neglecting to close an object may result in leaking native
resources. To debug this, set the environment variables
GEVENT_DEBUG=debug
and PYTHONTRACEMALLOC=n
before starting
the process.
The traditional cython-based libev backend will not leak if
close
is not called and will not produce warnings. The
CFFI-based libev backend will not currently leak but will produce
warnings. The CFFI-based libuv backend may leak and will produce
warnings.
Again, this is extremely experimental and all of it is subject to change.
See issue #790 for history and more in-depth discussion.
The C extension has been updated to use more modern Cython idioms and generate less code for simplicity, faster compilation and better cache usage. See pull request #1077.
Watcher objects may be slightly larger. On a 64-bit platform, a typical watcher may be 16 bytes (2 pointers) larger. This is offset by slight performance gains.
Cython is no longer preprocessed. Certain attributes that were
previously only defined in certain compilation modes (notably
LIBEV_EMBED) are now always defined, but will raise
AttributeError
or have a negative value when not available. In
general these attributes are not portable or documented and are
not implemented by libuv or the CFFI backend. See issue #1076.
Certain private helper functions (gevent_handle_error
, and part of
gevent_call
) are now implemented in Cython instead of C. This
reduces our reliance on internal undocumented implementation
details of Cython and Python that could change. See pull request #1080.
Next page: Changes for gevent 1.2