Also:
    - call shuffle correctly.
    - simplify test_suite()

input sets quickly, and will become much faster again soon (see TODO
in C function multiunion_m).

keys at the beginning of the BTree were deleted.

packtime > revid1's timestamp, in a way that's immune to differences
in platform clock resolution.  Solution given by Tim Peters.

the Standby storage version of the test has a hope of succeeding.
That storage ignores non-sensically out-of-sequence tids so as written
this test wouldn't have passed until after 3220-07-29 10:18:09.882353
<methuselah wink>.

We'll use a tid that we know is earlier than now.

storage needs to be more configurable).

blather log, and optionally replays the transactions into another
storage, timing the differences between the original and the new
storage spead.

objects), read_index would fail when trying to get the max of the
nonexistent oids in the transaction.

option.

    No need to generate code that is not needed!

the cache, their changes are flushed but they're not ghostified, so
their _p_changed == 0.  For objects not in the cache, they are
ghostified and so their _p_changed == None.  The tests had the check
values reversed.

the file limit is exceeded.

Add test case for long metadata.

The simplifications focus on avoiding tests that can be done
elsewhere and on avoiding excessive use of local variable aliases for
attributes.

Move tests for too-long transaction metadata to _begin(), where the
variables are set instead of here where they are read.

fixed ZClass-related type-checking breakage introduced in reformatting in rev 1.45. I guess there are no unit tests for ZClasses?

I think it's safe to run this code without extra checking.

Also, reformat a test in testTransaction.

constructor, since it is ignored and there is no current plan to
support two caches with almost-but-not-quite-the-same arguments.

This change has effects in many files.  The Connection and DB don't
pass this argument and don't bother setting it explicitly when it is
reset via DB APIs like setCacheDeactivateAfter().  XXX These APIs
remain, since existing code may depend on them, but they have no
effect.

New policy in cPersistence.c: A Persistent object can't have its
_p_jar set or deleted once it is in a cache.  Persistent already
implemented this policy for _p_oid; it seems safer to do the same for
the jar.

Add ringlen() method to cache objects (implemented as cc_ringlen).
This returns the length of the doubly linked list of non-ghost
objects.  Same as len(cache.lru_items()), but more efficient.  Only
used for testing at the moment.

In ring_corrupt(), don't raise a new exception if one has already been
set.  The old behavior masked useful information about the original
error / traceback.

Implement behavior for minimize() and full_sweep() as discussed on
zodb-dev.  minimize() ghostifies all unmodified objects.  full_sweep()
with age==0 is the same as minimize(), otherwise it's the same as
incrgc().

Reformat and/or reindent lots of code.

Use PyObject_Compare() instead of PyObject_Cmp() because it has a
simpler return value.

Fix a few more PyDict_SetItem() and PyDict_DelItem() calls to make
correct check for error return.

Break up the logic into a bunch of helper methods:
 _commit_objects()
 _commit_subtrans()
 _finish_one()
 _finish_rest()
and possibly
 _commit_error()

As a result of the changes, the high-level logic of a commit fits into
28 lines inside a try/finally.  There are lots of details hidden in
the methods, but the names capture the high-level behavior of the
helpers.

This fixes the transient failure in testExceptionInSubAbortSub() and
is presumed to be correct, too.  The old code had the comment:
    # This should never fail
but did not to guarantee that it never failed.  I'm assuming that the
write fix is to log the error and ignore it, because that's what the
call to tpc_abort() immediately above does.

XXX commit() is a really long method.  It would be nice to break it up
into easier-to-understand chunks.