What’s new in gevent 1.3#

Detailed information on what has changed is available in the Changes for 1.3. This document summarizes the most important changes since gevent 1.2.

gevent 1.3 is an important update for performance, debugging and monitoring, and platform support. It introduces an (optional) libuv loop implementation and supports PyPy on Windows. See Event Loop Implementations: libuv and libev for more.

Since gevent 1.2.2 there have been about 450 commits from a half-dozen contributors. Almost 100 pull requests and more than 100 issues have been closed.

Platform Support#

gevent 1.3 supports Python 2.7, 3.4, 3.5, 3.6 and 3.7 on the CPython (python.org) interpreter. It also supports PyPy2 5.8.0 and above (PyPy2 5.10 or higher is recommended) and PyPy3 5.10.0.

Caution

Python 2.7.8 and below (Python 2.7 without a modern ssl module), is no longer tested or supported. The support code remains in this release and gevent can be installed on such implementations, but such usage is not supported. Support for Python 2.7.8 will be removed in the next major version of gevent.

Note

PyPy is now supported on Windows with the libuv loop implementation.

Python 3.7 is in the process of release right now and gevent is tested with 3.7b4, the last scheduled beta for Python 3.7.

For ease of installation on Windows, OS X and Linux, gevent 1.3 is distributed as pre-compiled binary wheels, in addition to source code.

Note

On Linux, you’ll need to install gevent from source if you wish to use the libuv loop implementation. This is because the manylinux1 specification for the distributed wheels does not support libuv. The CFFI library must be installed at build time.

Greenlet Attributes#

Greenlet objects have gained some useful new attributes:

  • Greenlet.spawning_greenlet is the greenlet that created this greenlet. Since the parent of a greenlet is almost always gevent’s hub, this can be more useful to understand greenlet relationships.

  • Greenlet.spawn_tree_locals is a dictionary of values maintained through the spawn tree (i.e., all descendents of a particular greenlet based on spawning_greenlet). This is convenient to share values between a set of greenlets, for example, all those involved in processing a request.

  • Greenlet.spawning_stack is a frame -like object that captures where the greenlet was created and can be passed to traceback.print_stack().

  • Greenlet.minimal_ident is a small integer unique across all greenlets.

  • Greenlet.name is a string printed in the greenlet’s repr by default.

“Raw” greenlets created with spawn_raw default to having the spawning_greenlet and spawn_tree_locals.

This extra data is printed by the new gevent.util.print_run_info() function.

Performance#

gevent 1.3 uses Cython on CPython to compile several performance critical modules. As a result, overall performance is improved. Specifically, queues are up to 5 times faster, pools are 10-20% faster, and the gevent.local.local is up to 40 times faster. See pull request #1156, pull request #1155, pull request #1117 and pull request #1154.

Better Behaved Callbacks#

In gevent 1.2.2, event loop callbacks (including things like sleep(0)) would be run in sequence until we ran them all, or until we ran 10,000. Simply counting the number of callbacks could lead to no IO being serviced for an arbitrary, unbound, amount of time. To correct this, gevent 1.3 introduces gevent.getswitchinterval and will run callbacks for only (approximately) that amount of time before checking for IO. (This is similar to the way that Python 2 counted bytecode instructions between thread switches but Python 3 uses the more deterministic timer approach.) The hope is that this will result in “smoother” application behaviour and fewer pitfalls. See issue #1072 for more details.

Monitoring and Debugging#

Many of the new greenlet attributes are useful for monitoring and debugging gevent applications. gevent also now has the (optional) ability to monitor for greenlets that call blocking functions and stall the event loop and to periodically check if the application has exceeded a configured memory limit. See Monitoring and Debugging gevent Applications for more information.

New Pure-Python DNS Resolver#

The dnspython library is a new, pure-Python option for Name Resolution (DNS). Benchmarks show it to be faster than the existing c-ares resolver and it is also more stable on PyPy. The c-ares resolver may be deprecated and removed in the future.

API Additions#

Numerous APIs offer slightly expanded functionality in this version. Look for “changed in version 1.3” or “added in version 1.3” throughout the documentation for specifics.

A few changes of note:

  • The low-level watcher objects now have a close() method that must be called to promptly dispose of native (libev or libuv) resources.

  • gevent.monkey.patch_all defaults to patching Event.

  • gevent.subprocess.Popen accepts the same keyword arguments in Python 2 as it does in Python 3.

  • gevent.monkey.patch_all and the various individual patch functions, emit events as patching is being done. This can be used to extend the patching process for new modules. patch_all also passes all unknown keyword arguments to these events. See pull request #1169.

  • The module gevent.events contains the events that parts of gevent can emit. It will use zope.event if that is installed.

Library Updates#

One of the C libraries that are bundled with gevent have been updated. c-ares has been updated from 1.13.0 to 1.14.0 (c-ares release notes).

Compatibility#

This release is intended to be compatible with 1.2.x with no changes to client source code, so long as only non-deprecated and supported interfaces were used (as always, internal, non-documented implementation details may have changed). Here are some specific compatibility notes.

  • The resolvers have been refactored. As a result, gevent.ares, gevent.resolver_ares and gevent.resolver_thread have been deprecated. Choosing a resolver by alias (e.g., ‘thread’) in the GEVENT_RESOLVER environment variable continues to work as before.

  • The internal module gevent._threading was significantly refactored. As the name indicates this is an internal module not intended as part of the public API, but such uses have been observed.

  • The module gevent.wsgi was removed. Use gevent.pywsgi instead. gevent.wsgi was nothing but an alias for gevent.pywsgi since gevent 1.0a1 (2011).