processing the messages because the format might have changed a little bit.

(the tests themselves still aren't being run; we need decent
save/restore behavior from the logging package to do this properly)

logging directly

later

didn't recognize its version number.  As a result, 3.3a3 went out with
ZODB/__init__.py claiming it was version 3.3a2.  Repaired that, but by
changing what release.py looks for.

cause of the:

  TypeError: unpack() argument 2 must be string or read-only buffer, not None

failure often seen in overnight testrunner reports now, coming from the
checkConcurrentUpdatesInVersions test.  If the hypothesis is correct,
we'll see a failure of the new

        assert end is not None

instead.

returns a datetime.timedelta instead of a float.

Added test case and documentation.

thinks they live.

The old code used itertools.chain(), which didn't work as expected.
If the first of the two iterator grew after the chain had started
consuming the second iterator, the new element(s) in the first
iterator would never be consumed.

Fix this bug by eliminating the need for chaining.  Process all the
_added_during_commit objects at the end, each with its own
ObjectWriter instance.

Add explanation for __getattr__ and friends.

raised when an operation on a Connection can't be performed because the
Connection is in "a wrong state".

Connection.close() now raises this exception if the connection is
currently joined to a transaction (most obviously, if the transaction
the connection is joined to has modified objects waiting for commit
or abort).

Also changed other appropriate instances of RuntimeError to raise
ConnectionStateError instead (e.g., trying to load an object from a
closed connection).

Replace use of type('') with basestring.

The code didn't call setLocalTransaction(), so getTransaction() always
did the same thing as transaction.get().

essentially means a connection keep track of which objects from
the connection are modified, instead of transactions keeping track of that.
A good consequence hasn't yet been implemented:  if a connection is
closed with modifications still pending, we can detect that efficiently
now, and complain.