Difference between revisions of "Oslo/blueprints/time monotonic"
(→Rationale) |
|||
Line 7: | Line 7: | ||
Using time.time() to compute timeouts is wrong. The elapsed time may be wrong on update of the system clock. A monotonic clock should be used to avoid this issue. I propose to add a timeutils.time_monotonic() function. It will use time.monotonic() if available, or fallback to clock_getclock(CLOCK_MONOTONIC) on Linux. | Using time.time() to compute timeouts is wrong. The elapsed time may be wrong on update of the system clock. A monotonic clock should be used to avoid this issue. I propose to add a timeutils.time_monotonic() function. It will use time.monotonic() if available, or fallback to clock_getclock(CLOCK_MONOTONIC) on Linux. | ||
− | For a longer rationale, read the [http://legacy.python.org/dev/peps/pep-0418/ | + | For a longer rationale, read the [http://legacy.python.org/dev/peps/pep-0418/ PEP 418]. |
== Use time.monotonic() == | == Use time.monotonic() == | ||
− | Eventlet should use time_monotonic() instead of time.time(). Example of bug of time.time(): [https://www.mail-archive.com/openstack-dev@lists.openstack.org/msg18658.html | + | Eventlet should use time_monotonic() instead of time.time(). Example of bug of time.time(): [https://www.mail-archive.com/openstack-dev@lists.openstack.org/msg18658.html openstack-dev: time.sleep is affected by eventlet.monkey_patch()]. |
For greenlet, the monotonic clock can be used with something like that: | For greenlet, the monotonic clock can be used with something like that: | ||
Line 17: | Line 17: | ||
eventlet.hubs._threadlocal.hub = eventlet.hubs.get_default_hub().Hub(monotonic_time) | eventlet.hubs._threadlocal.hub = eventlet.hubs.get_default_hub().Hub(monotonic_time) | ||
− | time_monotonic() should be used for any function computing a timeout. Recent example: [https://review.openstack.org/#/c/71003/ | + | time_monotonic() should be used for any function computing a timeout. Recent example: [https://review.openstack.org/#/c/71003/ oslo.messaging: Add an optional timeout parameter to Listener.poll] |
== Implementation == | == Implementation == | ||
− | The following patch is based on the file time_monotonic.py from my Trollius project. Trollius is tested on Linux, Mac OS X, Windows, FreeBSD and OpenIndiana (which is almost Solaris): [https://review.openstack.org/#/c/85717/ | + | The following patch is based on the file time_monotonic.py from my Trollius project. Trollius is tested on Linux, Mac OS X, Windows, FreeBSD and OpenIndiana (which is almost Solaris): [https://review.openstack.org/#/c/85717/ timeutils: add time_monotonic() function]. The reimplement reuses time.monotonic() if available. |
== Alternatives == | == Alternatives == |
Latest revision as of 14:53, 7 April 2014
Authors
- Victor Stinner - victor.stinner@enovance.com
Rationale
Using time.time() to compute timeouts is wrong. The elapsed time may be wrong on update of the system clock. A monotonic clock should be used to avoid this issue. I propose to add a timeutils.time_monotonic() function. It will use time.monotonic() if available, or fallback to clock_getclock(CLOCK_MONOTONIC) on Linux.
For a longer rationale, read the PEP 418.
Use time.monotonic()
Eventlet should use time_monotonic() instead of time.time(). Example of bug of time.time(): openstack-dev: time.sleep is affected by eventlet.monkey_patch().
For greenlet, the monotonic clock can be used with something like that:
eventlet.hubs._threadlocal.hub = eventlet.hubs.get_default_hub().Hub(monotonic_time)
time_monotonic() should be used for any function computing a timeout. Recent example: oslo.messaging: Add an optional timeout parameter to Listener.poll
Implementation
The following patch is based on the file time_monotonic.py from my Trollius project. Trollius is tested on Linux, Mac OS X, Windows, FreeBSD and OpenIndiana (which is almost Solaris): timeutils: add time_monotonic() function. The reimplement reuses time.monotonic() if available.
Alternatives
- https://pypi.python.org/pypi/Monotime
- On Linux, it uses CLOCK_MONOTONIC_RAW which is wrong according to my PEP 418. Extract: "Slewing is generally desirable (i.e. we should use CLOCK_MONOTONIC, not CLOCK_MONOTONIC_RAW) if one wishes to measure "real" time (and not a time-like object like CPU cycles)."
- Monotime module doesn't support Windows nor Solaris. Does it matters for Oslo Incubator?
- It doesn't reuse time.monotonic() if available.
- python-monotonic-time
- It doesn't look to be available on PyPI: https://github.com/gavinbeatty/python-monotonic-time