The trunk now has the same code ZEO 1.0b5 plus a few minor changes.

acceptability of serials=None for the object-creation store() call.
Note that _dostore() itself never exercises this condition because it
coerces its own serial argument to ZERO (a.k.a. '\0'*8) when None is
passed in.

The Mapping_update() method iterated over a sequence of 2-tuples, but
held an extra reference to the sequence.  The sequence object was
DECREFed on an error exit but not on the normal return.

The fix reworks the control flow of the function to make the reference
counting behavior a little clearer:

   - Test for call with no arguments and return immediately if so.

   - If the argument is a sequence, don't INCREF it.

   - If the argument passed isn't a sequence, call its items() method
     and store the result in the local variables seq and items.

   - On exit, do Py_XDECREF(items).  This will do a decref only if a
     the second case above applied.

Two other minor nits:

   - Make consistent use of spaces around = in assignments (as opposed
     to a fix of both)

   - Remove assignment from conditional expression.

pack() uses a coarse-grain timer to decide which revisions to look at
during a pack.  The tests depend on a pack() called after a commit to
see that committed transaction, but the implementation can't guarantee
that it will.

In particular, pack() converts its time argument to a TimeStamp by
truncating to the nearest second.  If a transaction T1 commits and pack
is called (with the current time as its argument) during the same second,
the pack() will not consider T1.  This was causing intermittent, but
frequent test failures on Windows.

The fix is to add a small delay in the pack test cases to guarantee that
pack(time.time()) will always see the last committed transaction.

If the packpos returned by the read_index() call is 4, return immediately.
In this case, the very first transaction, which contains the root object,
is after the pack time.  This is a degenerate case, but it seems clearest
to do nothing.

This change eliminates an unexpected traceback in pack() that occurred
frequently, but not every time, when testing on Windows.  When
read_index() returned 4, a seek() call in _redundant_pack() would fail
with a bad argument.

Add an assertion to _redundant_pack() because it should never be called
with bogus arguments as a result of the other change.

Don't test for explicit end-of-line markers, since they vary across
platforms.  Instead, strip() the line read and test for the actual content.

If a storage doesn't support T.U. and is old enough, you need to first
check if it has a supportsTransactionalUndo attribute.

There are a few problems with Zope 2.3 that prevented the tests from
succeeding.  Since the 1.0 release must support Zope 2.3, we'll need
to hack the various tests to work around them.

    Assuming releases after 1.0 don't need to work with 2.3: After the
    1.0 final release, I'll tag both ZEO and ZODB trees and then undo
    all these silly changes.

Problems:

    - no support for transactional undo or getSerial()
    - FileStorage commitVersion() and abortVersion() didn't check for
      bogus arguments

This implementation is only faster for an object that is not modified
in a version.  For an object modified in a version, it should be about
the same as the default getSerial() in BaseStorage.

situation.

Didn't write a more interesting test in checkAbortVersion:

        # XXX Not sure I can write a test for getSerial() in the
        # presence of aborted versions, because FileStorage and
        # Berkeley storage give a different answer. I think Berkeley
        # is right and FS is wrong.

BaseStorage implements this in terms of modifiedInVersion() and
load().  Other storages may (perhaps even should) implement it more
efficiently.

now a full success :-)

(Minor change to tb handling to avoid unnecessary binding to None for
v and tb.)

Prevents memory leak for each _restore_index() call.

exceptions (this error occured when trying to access an
OOBTree with a string with non-ascii encoding e.g. latin1.
In this case Python raises a Unicode exception that has not
been caught and not propagated to the application.

Replace explicit equality tests in assert statements, which aren't
executed when running python -O, when self.assertEquals(), which also
has the side effect of producing better error messages when the test
fails.

  o Land unit test cleanup from branch.

  o Known breakage as of 2001/10/17:

    - TAL test suite is very broken (TAL is currently suppressed in
      lib/python/.testinfo).

    - AccessControl tests don't run if run first;  lib/python/.testinfo
      pushes them to the end of the queue, at which point they *do* run.
      Global side effects suck!

    - lib/python/BTrees/testBTreesUnicode.py:  weird flakiness around
      comparison of ASCII and Unicode strings;  breaks in different
      ways under Python 2.1 and 2.2.

    - lib/python/Products/PageTemplates/tests:  three tests break due
      to security asserctions.

strings as keys

Add getLogFile() method that provides hook for SpreadLog subclass.

This change includes a minor change to semantics for
reinitialization().

There are three chief performance-related changes:

    - Do not check whether logging is enabled in the actual log call.
      Instead check enviroment variables when the module is imported
      and whenever an explicit initialize() call is made.

      This might seem like it limits the user's ability to turn
      logging on and off, but I don't believe it does.  Logging is
      controlled by environment variables.  Once the Python script is
      working, the environment variables won't change.

      This change makes the case of no logging fast, instead of slow.

    - Each log call goes through a method on the stupid_log_write()
      object instead of through __call__().  It's much faster to call
      a method than an instance.

    - Use "print >> file" rather than "file.write()".  It appears to
      be a bit faster.

Replaced _set_stupid_dest() with _set_log_dest().

Also, reformat code using guidlines from the Friends of Whitespace.

Fix newTimeStamp() so that it always returns a timestamp.
XXX still not sure that it works.

Replace long multiplies with long shifts in U64() and u64().

    - use update() method where appropriate
    - don't pass None as second argument to get(). it's the default.
    - consistent space after comma between args

In tpc_vote() rename local variable id to tid.

Add doc string to read_index(), helpful but not complete yet.

In read_index(), replace an iteration over the tindex for each
transaction with a few calls that do the work with much less
overhead.