when a transaction is about to be written. It would be bad if we
"flipped" the cache twice during a transaction, as we would have
incomplete transaction data.  Note that this means that we could
exceed the target cache size if the data for a transaction exceeded
the target size (/2).

Changed the way that the storage is closed. Our close method may be
called from a separate thread than the async main loop. We only want
the async main loop to close our _call (rpc) object. Instead of
closing it directly, we call closeIntensionally(), which, in return,
requests that the async main loop closes it, via the select trigger.

Added seek before write in invalidate method. Without this, strange
things happen on NT and solaris. (I was probably breaking a rule of
thumb that says never to do a write() right after a read() even though
tell() would indicate I was in the right place.)

Added logic to update the cache index on an invalidate.

Added a separate call to check the cache size so that we only check
when a transaction is about to be written. It would be bad if we
"flipped" the cache twice during a transaction, as we would have
incomplete transaction data.  Note that this means that we could
exceed the target cache size if the data for a transaction exceeded
the target size (/2).

  - Serving multiple storages,

  - Serving storages other than file storages.

access.

The protocol was changed on the client, but not the server.
The client has been changed back. This part of the protocol
will go away in r0.3.

(e.g. when moving data between a DCOracleStorage and a FileStorage),
especially the method copyTransactionsFrom, as in:

  import ZODB.FileStorage, DCOracle, DCOracleStorage

  # Copy data.fs to oracle
  DCOracleStorage.Full(
    lambda : DCOracle.Connect('user/pw@sid')
    ).copyTransactionsFrom(
      ZODB.FileStorage.FileStorage('Data.fs', read_only=1)
    )

any current storages.

included unless the -L option is provided.

a transaction id to be spacified, rather than being defined
by the storage. This should *only* be used when copying data
from one type of storage to another.

from one storage to another.

than the (timestamp of the first record written after the) last pack.
Packing to a time earlier than the last pack can cause serious data
lossage.

requested disregarded the filter. That is, the arguments first and
last should be applied *after* filtering.

database rather than the mounted database.

attributes are registered.

got registered. (Deja vu all over again. ;)

Changed to commit objects in registration order and to allow
additional objects to get registered while commiting. (The
later was wanted by PJE.)

Added a comment in commit to explain the "invariant".

Fixed bug that caused objects to get multiply aborted.  Now always
clear object registry on commit and abort. This seems like a good
idea anyway. (We were relying on free_transaction to have the
same effect. This faild miserably in single-threaded apps and
in certain error conditions.

Added a teeny optimization in get_transaction().

transparent Pythonic sematics for __setattr__ and __delattr__.
If these methods are implemented, then they *must*::

  self._p_changed=1 # prefered

or::

  self.__changed__(1) # if people get the heebe-jeebees from setting attrs

to manually register that an object's persistent state has changed, if,
in fact, it has.

Fixed "while objects" loop in commit to do the right thing with the object list element (toss it) if it doesn't have a _p_jar.  Not doing this was causing an infinite loop.