Irony:  While these are the only accurate docs that
exist, FileStorage doesn't actually implement it this
way.  Fixing that is the next step.

releases" -- had become impossible to follow.

There are no tests for this.

If a threaded transaction manager ever passed None to
the Transaction constructor's `synchronizers` argument, then
synchronizers registered later in the same transaction
were invisible to the transaction, and so their afterCompletion()
methods wouldn't get called when the transaction ended.

This is effectively a beta release, but I don't want to
take the time to do a full-blown release right now.

whenever TransactionManager.begin() is called.

Connection implements that, and changes its ISynchronizer
afterCompletion() method, to call sync() on its storage
(if the storage has such a method), and to process
invalidations in any case.

The bottom line is that storage sync() will get done "by
magic" now after top-level commit() and abort(), and after
explicit TransactionManager.begin().  This should make it
possible to deprecate Connection.sync(), although I'm not
doing that yet.  Made a small but meaningful start by
purging many sync() calls from some of the nastiest ZEO
tests -- and they still work fine.

Added more comments.  Changed some comments to English.
Renamed some attributes and methods for clarity.
Switched to using a Python WeakKeyDictionary instead of
rolling our own out of a Python dict and raw weakrefs.

Added new test checkSubtxnCommitDoesntGetInvalidations to
verify that a longstanding bug in subtransaction commit is
repaired.

Jim (Fulton) discovered this in ZODB 3.4's code, while implementing
savepoint/rollback.  Same bugs had been there at least since ZODB 3.1.

Also added news about the bug.

for savepoint management are:

- All savepoints for a transaction should be invalidated when the
  transaction commits or aborts

- If a savepoint is rolled back, then all savepoints after it within 
  a transaction must be invalidated.

We previously implemented these requirements by organizing transaction
savepoints into a doubly linked list.  This was overkill.  We didn't
have need for such fine-grained ordering.  This strategy had the
disadvantage that it kept all savepoints around until the transaction
ended.  Savepoints could be expensive to keep and it's possible that
some applications could keep a lot of them.

The new stragey is to:

- Keep weak references to savepoints.  We can forget about savepoints
  that the application isn't using.  Any resources used by these
  savepoints can be freed.

(We have to keep a strong reference to the last savepoint used for
  a subtransaction.)

- We assign indexes to savepoints within a transaction.  When a
  savepoint is rolled back, in addition to invalidating that
  savepoint, we also invalidate savepoints with a higher index.

A side effect of this change is that code using the savepoint API
should interfere less with code using subtransactions.  Of course, we
really need to phase out code that uses subtransactions.

It is likely that we can leverage this change in strategy to speed
creation of ZODB connection savepoints.  Creating a ZODB connection
savepoint now requires copying the savepoint storage index.  This
index could become large.  If applications aren't holding on to old
savepoints, then it is possible that we could avoid this copy.

returns an iterable producing the registered beforeCommit hooks.

just the savepoint being rolled back because later savepoints
might involved data managers that hadn't joined when the savepoint
being rolled back was created.

Now, when a data manager joins and we have savepoints, we create a
data manager savepoint for the new data manager and add the
datamanager savepoint to all previous transaction savepoints.  Note
that this data manager savepoint can be a special savepoint that just
calls abort on the data manager when it is rolled back.

Removed the freeme argument.  This argument is part of the
implementation, not the public interface.

Removed the subtransaction argument.  Although it is still supported,
it isn't part of the pblic interface.

Added missing documentation of the savepoint method.

Added demonstration of transaction non-commitability after savepoint
or savepoint rollback failure.

Updated "previous commit failed" error to "previous operation failed".

transaction savepoints.

Updated some interface declarations.

Once we start using savepoints, we need to make sure that all data are
committed through the savepoints. Otherwise, things can get committed
in the wrong order, leading to conflicts.

savepoint, then modified and then aborted.

Also added missing code to clear registered objects when a savepoint
was rolled back.

(And also added interfaces and rearranged some code to hopefully make
it easier to read.)

Also fiddled things to make this look like release 3.4a4.

In addition, repaired the useless default logging in runzeo.py.

Here's Sidnei's Zope 2.7 checkin comment:

* Borrow Zope's 'Signal' mechanism for Windows, if available, to
  implement clean shutdown and log rotation handlers for Windows.

* Back to creating a .PID for ZEO, so external programs that wish to set
  the 'signal' can get the PID and therefore derive the signal name.
  Currently only necessary on Windows but created on all platforms which
  implement os.getpid(), as long as the 'pid-filename' option is set,
  or the 'INSTANCE_HOME' environment variable can be found.

DemoStorage:  Added implementations for registerDB() and new_oid().  As
Tres discovered the hard way, wrapping a ZEO client storage as a
DemoStorage base storage yields insane behavior otherwise.

BaseStorage.new_oid():  Rewrite to eliminate recursion, and hence also the
need for the undocumented and irregular `last=` argument.

Other:  removed the `last=` argument to new_oid() every place that felt
compelled to spread that insanity ;-).  Seriously, it served no purpose
at all elsewhere, and looks like people just cut 'n pasted in fear.