- Restart or pack Zope, causing an index to get written

  - Undo some transactions

  - Shutdown Zope ungracefully (e.g. kill)

When Zope comes up, the index points to undone records for some
objects. This can lead to all sorts of problems. :(

To overcome this problem, we now remove the index after an
undo.

This increases the likelyhood that it will take longer to restart
next time, but reduces the chance of a bad index.

Renamed some classes to make their intent clearer.

Fixed a silly typo in zrpc.

traceback info rectified by the last, more substantive checkin...

object that has previous records.  This caused a panic on startup
because the previous record pointer did not agree with the system's
knowledge that there was, indeed, a previous record. This is a
pretty benign situation, so, I've changed the check to log a warning
of the previous record pointer is zero and there is a previous record.

instances as non-persistent.

got registered.

.__class__.   This caused acquisition wrappers to get pickled. :(

ignored and an indeterminate value used instead.

a transaction could, in rare cases cause inconsistent computations.

Also added argument to incrgc to allow more aggressive GC.

storage errors.

to get version cache settings.

when their state was loaded. :-[

moving parts.

The minimize machinery didn;t use the correct time units.

The incremental GC had some bugs that made it too conservative.

the actual caches.

to fail.

Fixed memory leak caused by incomplete PER_DEL.

out of memory can force aggressive cache behavior:

  cacheGC() -- do an incremental GC

  cacheFullSweep([dt]) -- Make a pass through objects in the cache
     removing unreferenced objects and, if dt > 0, deactivating
     objects accessed more than dt seconds ago.

  cacheMinimize([dt]) -- Make multiple passes through objects in the cache
     removing unreferenced objects and, if dt > 0, deactivating
     objects accessed more than dt seconds ago. Keep making passes
     until the cache size doesn;t change.

Added a coptimizations module.  The first optimization is
to provide an implementation of persistent_id in C.  This
routine, which is called extremely often while pickling was
found to be a significant bottleneck.

Fix a bug in object deactivation.
Overridden versions of _p_deactivate were not called.
This meant that BTrees were never deactivated.