If transactions are supported - reverse the order of action in creating
placeholder - create a record in Global_Fields first, then add fields
and roll-back if unsuccessful.

===============

BUG (am, 3/31/2003): If we're searching on:

 [ 'Sub/name','cs','test' ],
 'or',
 [ 'title','cs','test' ]

and there are no entries at all in Sub for some element where 'title'
matches 'test' that element won't be shown. We need outer join or
something similar to deal with it. We need to track if a specific
sub-table is required or not and join accordingly. Looks damn
complicated if I'm right.

===============

Multiple read-servers, read_pool argument to odb. For better chance
of returning in-sync data it is probably worth keeping track of last
update issued to the main server and querying the main server for a
couple of seconds after each update before switching to normal read_pool
selection.

Also, it is probably better to only keep two connections -- to master
server and to a random read_pool server. If read server goes away -
reconnect to another read_pool server. It has no merit to keep cycling
through all read pool servers continuously. Or may-be we should have a
limit on how many connections to read servers to keep/cycle through.

===============

(DONE) Multivalue put.

===============

Make sure that there is a glue() method available on both Lists and
Hashes and that fetch() also works on all of them.

===============

Polymorphism. See letter from Jean-Michel Hiver for details and ideas.

===============

Optimization for keys() and values() called in scalar context to issue
count(*) instead of returning actual array. Probably a special method
like count(search_args) too for more complex conditions.

===============

(!) Locking and/or transactions support. Safe adding of a new item with
pre-defined ID using checks that no such item already exists. Somehow
must deal with race condition if two attempts to add the same item are
made at the same time.

===============

Fix destroy() on hash objects. It currently does not remove the object,
but empties it incorrectly.

===============

Make it possible to search on the same deep object or on any. Syntax is
similar to:
 [ [ 'Interests/1/interest_id', 'eq', 'qwerty' ], 'and',
   [ 'Interests/1/level', 'eq', 123 ],
 ], 'and',
 [ 'Interests/2/interest_id', 'eq', 'asdf' ]

Meaning: look for all interests where interest_id is 'qwerty' and level
is 123 on the same object and where there is also an interest with
interest_id equal to 'asdf' in the same base object.

Also:
 [ 'Interests/*/interest_id', 'eq', 'asdf' ], 'and',
 [ 'Interests/*/interest_id', 'eq', 'qwerty' ]
to denote matching on any object. The difference is merely to add the
same sub-table again as another letter or not. Probably change query
syntax to use 'left join' instead of explicit index comparisions.

===============

Negation of 'wq' and 'ws' in search() (Marcos' idea)

===============

fill() method or rather extended put().

===============

temporary_data() method to get per-object space to store something
temporary that would not be stored in the object on disk. Some cached
values or something like that. Just a hash or SimpleHash reference in
fact.

===============

(DONE) Transactions, InnoDB support. trans_begin(), trans_rollback() and
trans_commit().