nova/doc/source/threading.rst

Threading model

All OpenStack services use green thread model of threading, implemented through using the Python eventlet and greenlet libraries.

Green threads use a cooperative model of threading: thread context switches can only occur when specific eventlet or greenlet library calls are made (e.g., sleep, certain I/O calls). From the operating system's point of view, each OpenStack service runs in a single thread.

The use of green threads reduces the likelihood of race conditions, but does not completely eliminate them. In some cases, you may need to use the @lockutils.synchronized(...) decorator to avoid races.

In addition, since there is only one operating system thread, a call that blocks that main thread will block the entire process.

Yielding the thread in long-running tasks

If a code path takes a long time to execute and does not contain any methods that trigger an eventlet context switch, the long-running thread will block any pending threads.

This scenario can be avoided by adding calls to the eventlet sleep method in the long-running code path. The sleep call will trigger a context switch if there are pending threads, and using an argument of 0 will avoid introducing delays in the case that there is only a single green thread:

from eventlet import greenthread
...
greenthread.sleep(0)

In current code, time.sleep(0）does the same thing as greenthread.sleep(0) if time module is patched through eventlet.monkey_patch(). To be explicit, we recommend contributors use greenthread.sleep() instead of time.sleep().

MySQL access and eventlet

Queries to the MySQL database will block the main thread of a service. This is because OpenStack services use an external C library for accessing the MySQL database. Since eventlet cannot use monkey-patching to intercept blocking calls in a C library, the resulting database query blocks the thread.

The Diablo release contained a thread-pooling implementation that did not block, but this implementation resulted in a bug and was removed.

See this mailing list thread for a discussion of this issue, including a discussion of the impact on performance.