9/22/04: Moved code into Subversion repository. Changes now managed in SVN log.

1.2.3 (9/3/04):
    1. Decoupled construction of stores from arena._load() into add_store().
    2. Exceptions in SocketServer now get passed back to the calling app
        and reraised.
    3. Fixed nasty latent bug where engine FILTER rules weren't filtering
        if the set already had elements; it would just pass through every
        existing element without evaluating them against the filter operand.
    4. Wrote dejavu.analysis.py, which provides aggregation and analysis
        tools to dejavu applications. Moved dejavu.unit_sort to analysis.sort.
    5. storeado.CALL_FUNCTION failed on some logic functions. If the func
        object was stored in co_consts, it would error on .endswith. Now
        falls through, setting imperfect, so dejavu can test with the lambda.

1.2.2 (8/18/04):
    1. Changed storage/sockets to use blocking sockets, resulting in a 40% speedup.
    2. Use faster CPickle if possible in sockets, storeado, engines.
    3. Moved on_forget calls after the unit is forgotten. All known use
        cases are OK with this; some needed it to update parent sums.

1.2.1 (8/10/04):
    1. Got rid of unit.has(), which didn't fit the new unit formation process.
    2. Rewrote CachingProxy.sweep() to destroy temporary units (which rolls
        back the reserve(), in effect).
    3. CachingProxy.shutdown duplicated sweep_all().
    4. Sandbox was setting Unit.dirty = False on recall; that should be left
        to SM's, since they might cache Units.
    5. storeado.AdapterToADOFields.coerce_datetime() didn't trap
        value == None (tried to add None + None).

1.2.0 (8/9/04):
    1. Sandbox.forget had a NameError: store = self.arena.storage(cls)
    2. Storage.CachingProxy.sweep errored on some _recallTimes[id]
    3. for job, project in recall(Job, [expr, [Project, [expr]]]).
        Recalls associated Units in a single SM query.
    4. Added a special case for Expression(lambda x: x.ID == CONST),
        to save the database hit if it's found in the cache.
    5. Profiling: Changed storeado.py, storage/sockets.py to set
        unit._properties directly to save __set__ overhead (which isn't
        really needed until the unit is concrete).
    7. Profiling: added unit._property_types for faster lookup.
    8. After profiling, added _co_code_attrs to codewalk because dir() was so slow.

1.2.0 RC (8/5/04):
    1. Did another big rework. Sandboxes now use straight dicts for caching
        Units, and are never shared across requests. Any sharing has been
        pushed into subclasses of StorageManager, for example, CachingProxy.
        This should also give us the framework to write multi-Unit recall.
    2. Given the new chainable StorageManagers, added "Load Order" and
        "Shutdown Order" to config options.
    3. Merged config "Cache" sections with "StorageManager" sections (since
        there are no more Cache classes).
    4. Changed arena.roster to be a Prism; removed RegisteredUnitClass class.

1.1.3 (7/30/04):
    1. Dang. Forgot to have memorize() set dirty with all the refactoring.
    2. Times and datetimes not being saved correctly (math/parens typo).
    3. Times and datetimes not being retrieved correctly due to naive
        rounding of times, which are fractional.

1.1.2 (7/29/04):
    1. Change storeado to drop UnitCollection tables instead of emptying them;
        this makes modifying the type of the ID column easier.
    2. engines.take_snapshot(): nontemp Engines now create temp snapshots.
    3. UnitEngine.on_repress removed--it isn't necessary.

1.1.1 (7/28/04):
    1. storeado was failing on load_collection when looking up class.
    2. storeado wasn't retrieving currencies correctly (lost fractions).

1.1.0 (7/27/04):
    1. storeado.load_collection had a bug setting idtype.
    2. Added shutdown() to Storage Managers.

1.1.0 RC (7/26/04):
    1. distinct() functions at sandbox, cache, storage levels.
    2. Completely reworked caching mechanism so apps could flush caches
        after each request (which turns out to be most efficient for
        web apps, rather than having each thread serialize access to
        shared caches). UnitServer is gone, and Unit caches are now owned
        by Sandboxes, which are doled out to apps.
    3. Sandbox attribute of Units gets set on memorize, not init. No more
        Unit.memorize(); instead, it's sandbox.memorize(unit), and any
        custom unit code that was in overridden Unit.memorize functions
        now needs to be in Unit.on_memorize().
    4. storeado.save_collection: Unit.ID.type was not being detected
        correctly; hence no save of UnitCollections.

1.0.3 (7/21/04):
    1. Since cache wasn't being indexed anyway, went back to a plain dict.
    2. Added an option to storeado to create a new ADO Connection
        for each thread. Doesn't seem to make a difference on speed.
    3. Added UnitServerWeakref.
    4. Units were being saved on repress() even if not dirty.
    5. Changed default UnitServer to no cache. UnitServerCaching should
        be considered broken until its cache management can be made
        thread-safe (that is, it's far too slow because it has to
        synchronize every recall).
    6. Modified Arena to handle unit class->server mappings more
        statically, with a 'roster' dictionary. ALL unit classes must
        now be registered via arena.register() or .associate().
    7. Made Unit.server an attribute, set at __init__, rather than
        a method. Consequently, was able to remove some of the litmus
        testing done in UnitServer methods--the litmus should be tested
        long before those calls; i.e., when arena.roster is populated.

1.0.2 (7/19/04):
    1. Reworked recall by removing Recallers completely. Big speedup.

1.0.1 (7/16/04):
    1. Added the COPY rule to engines.py.

codewalk.py 0.3.1 (4/29/04):
    1. Added KeywordInspector.
    2. Made LambdaDecompiler use function __name__, not repr() (early
        binding makes e.g. datetime.date into <type datetime.date>).
    3. Added safe_tuple.
    4. Added TaintableStack.maxsize to help stack calculation.
    5. Forgot to call deref_cell from visit_LOAD_DEREF! D'oh!
    6. Put visit_ methods in alpha order.
    7. Added name_index to Rewriter.
    8. Got rid of extra args to EarlyBinder.
    9. Added reduction of getattr to EarlyBinder as an option.
    10. Combined 'ands' and 'ors' into 'targets' in LambdaDecompiler.
    11. Fixed parenthesis problems in LambdaDecompiler.
    12. Added BranchTracker class.

Prior to July 2004, Dejavu was integrated with a Python application
framework, which is now called "Cation". Development on the
previously-integrated project began in September 2003.