A thread local variable is a variable that appears to be a "global" variable to an application which uses it. However, unlike a true global variable, one thread or process serving the application may receive a different value than another thread or process when that variable is "thread local".
When a request is processed, Pyramid makes two thread local variables available to the application: a "registry" and a "request".
How are thread locals beneficial to Pyramid and application developers who use Pyramid? Well, usually they're decidedly not. Using a global or a thread local variable in any application usually makes it a lot harder to understand for a casual reader. Use of a thread local or a global is usually just a way to avoid passing some value around between functions, which is itself usually a very bad idea, at least if code readability counts as an important concern.
For historical reasons, however, thread local variables are indeed consulted by various Pyramid API functions. For example, the implementation of the pyramid.security function named authenticated_userid() (deprecated as of 1.5) retrieves the thread local application registry as a matter of course to find an authentication policy. It uses the pyramid.threadlocal.get_current_registry() function to retrieve the application registry, from which it looks up the authentication policy; it then uses the authentication policy to retrieve the authenticated user id. This is how Pyramid allows arbitrary authentication policies to be "plugged in".
When they need to do so, Pyramid internals use two API functions to retrieve the request and application registry: get_current_request() and get_current_registry(). The former returns the "current" request; the latter returns the "current" registry. Both get_current_* functions retrieve an object from a thread-local data structure. These API functions are documented in pyramid.threadlocal.
These values are thread locals rather than true globals because one Python process may be handling multiple simultaneous requests or even multiple Pyramid applications. If they were true globals, Pyramid could not handle multiple simultaneous requests or allow more than one Pyramid application instance to exist in a single Python process.
Because one Pyramid application is permitted to call another Pyramid application from its own view code (perhaps as a WSGI app with help from the pyramid.wsgi.wsgiapp2() decorator), these variables are managed in a stack during normal system operations. The stack instance itself is a threading.local.
During normal operations, the thread locals stack is managed by a Router object. At the beginning of a request, the Router pushes the application's registry and the request on to the stack. At the end of a request, the stack is popped. The topmost request and registry on the stack are considered "current". Therefore, when the system is operating normally, the very definition of "current" is defined entirely by the behavior of a pyramid Router.
However, during unit testing, no Router code is ever invoked, and the definition of "current" is defined by the boundary between calls to the pyramid.config.Configurator.begin() and pyramid.config.Configurator.end() methods (or between calls to the pyramid.testing.setUp() and pyramid.testing.tearDown() functions). These functions push and pop the threadlocal stack when the system is under test. See Test Set Up and Tear Down for the definitions of these functions.
Scripts which use Pyramid machinery but never actually start a WSGI server or receive requests via HTTP such as scripts which use the pyramid.scripting API will never cause any Router code to be executed. However, the pyramid.scripting APIs also push some values on to the thread locals stack as a matter of course. Such scripts should expect the get_current_request() function to always return None, and should expect the get_current_registry() function to return exactly the same application registry for every request.
You probably should almost never use the get_current_request() or get_current_registry() functions, except perhaps in tests. In particular, it's almost always a mistake to use get_current_request or get_current_registry in application code because its usage makes it possible to write code that can be neither easily tested nor scripted. Inappropriate usage is defined as follows:
Use of the get_current_request() function in application code is still useful in very limited circumstances. As a rule of thumb, usage of get_current_request is useful within code which is meant to eventually be removed. For instance, you may find yourself wanting to deprecate some API that expects to be passed a request object in favor of one that does not expect to be passed a request object. But you need to keep implementations of the old API working for some period of time while you deprecate the older API. So you write a "facade" implementation of the new API which calls into the code which implements the older API. Since the new API does not require the request, your facade implementation doesn't have local access to the request when it needs to pass it into the older API implementation. After some period of time, the older implementation code is disused and the hack that uses get_current_request is removed. This would be an appropriate place to use the get_current_request.
Use of the get_current_registry() function should be limited to testing scenarios. The registry made current by use of the pyramid.config.Configurator.begin() method during a test (or via pyramid.testing.setUp()) when you do not pass one in is available to you via this API.