New submission from Jean-Paul Calderone <exar...@divmod.com>: When a frame's f_locals attribute is accessed, one of two things happens:
* If this is the first access, a new dictionary is created and it is populated with the locals from the frame (since they are not stored internally as a dictionary). The dictionary is both saved and returned. * Otherwise, the previously saved dictionary is updated to reflect the current locals of the frame and then returned. This means that the lifetime of locals in the frame is strangely altered if the frame's f_locals attribute is ever accessed. In particular, the locals are kept alive until the frame object is discarded or the f_locals attribute is accessed again. The drawback of this is that references to locals may be "leaked" in a surprising and difficult to understand way. Another is that when debugging an application, this will cause objects to have a different lifetime than they will when the application is run without the debugger. One way the former of these issues has manifest is described in this Twisted bug report: http://twistedmatrix.com/trac/ticket/3853 There, I suggested three possible solutions, in the form of a replacement API for the f_locals attribute which either: * Gave you a copy of the locals at the current time and didn't pretend it would be automatically updated * Gave you a view onto the locals at the current time, instead of a real dictionary, and didn't actually hold any references to the locals itself, only granted access to the "canonical" locals data in the frame. * Gave you a real dictionary of the locals but didn't hold a strong reference to that dictionary internally, so that if all external users of the dictionary went away, the frame itself would also stop referring to the locals. ---------- components: Interpreter Core messages: 88365 nosy: exarkun severity: normal status: open title: frame.f_locals keeps references to things for too long type: behavior _______________________________________ Python tracker <rep...@bugs.python.org> <http://bugs.python.org/issue6116> _______________________________________ _______________________________________________ Python-bugs-list mailing list Unsubscribe: http://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com
