The cache is intended to keep weak references to objects.  An object
should be deallocated when the only references to it are in the
cache.  If an object is unghostified and added to the list of live
objects, the list should have a borrowed reference.  Otherwise, the
cache will keep the object alive after all other references have
disappeared.

time.  Instead use std PyDict_GetItem() and incref only when
necessary.

The cache decrefed it in percachedel.

PyDict_DelItem() will undo the other INCREF.

Decref its reference to the cache.
In a debug build, account for the two bogus additions to _Py_RefTotal.

Also, reindent initcPickleCache().

Remove object_from_oid() which appeared to be mis-used about half the
time.  Instead use std PyDict_GetItem() and incref only when
necessary.

Make sure an object's reference to its cache is decref'd when the
object is deallocated.  This change seems to eliminate leaking pickle
caches.

Remaining mystery: Why do objects stay in the cache even when there
are no references to them?

release.

descriptions.

and are of questionable value.  "constructor" is also generally unsafe;
applications that need it can provide their own version of this if necessary.

- simplify document structure at one point

- fixed datatype of the values for "handler" attributes
- removed/commented out all mentions of schema component extension;
  that's really not complete and should probably be thought through
  again
- moved docs for legacy API out; these should not be generally
  documented
- misc. small fixes

The manager should raise an exception in its prepare() method rather
than returning a boolean to indicate failure.  Rationale: The txn
manager can't raise a reasonable exception, because it doesn't know
what the data manager couldn't prepare.

It may be reasonable to change this more, but the benefits aren't clear.
(These changes will be used from the new-install-branch.)

make it easier to re-locate an instance.

Don't ever stop after the first object.  It leaves all the other
objects in the version unmodified.

Fix edge case in pack() -- backpointer to object created in version.

pair) from the bucket into the set-iteration struct twice.  I don't
believe this had any visible effect, it was simply pointless and wasted
a little time (discovered by eyeball).

Other than fixes of nonconforming whitespace, just a couple of $Id$
docstrings has been added.

It's not clear what the right answer is here.  The implementation
aborts any remaining resource managers and hopes for the best, which
isn't very satisfying.  Perhaps the hosed feature should come back.

Add some comments about the issues involved.  Revise the tests to
cover the current behavior.

XXX Should the manager duplicate these checks?

Make txn_factory an attribute of the base class.

Raise an exception when prepare() returns False, rather than
automatically aborting.

Pass transaction object to Rollback() so that rollback() method can
check state of transaction.

Add IllegalStateError calls to prevent assertions from failing.
XXX Should the manager duplicate these checks?

Add suspend() and resume() to non-threaded txn manager.

Fix bug that caused threaded suspend() to fail with KeyError for
thread with no current transaction.

classes.