And the next round, this time I am closer to Durus.

   If I have a deep hierarchy a.b.c.d, and fetch the object "a", pickle
(and hence Durus and ZODB) bring me all the objects. If I need a lazy
attribute that does not fetch the referenced object until I specifically
say "a.b" I have to implement it myself using properties and store an oid
or other index of "b", right?
   Are btrees, PersistentDicts, PersistentLists lazy it that sense? Do they
fetch all objects at once or fetch them one by one at need?

How can I serialize parallel writes? Two processes want to do
   btree = root[btree_name]
   btree[key] = value
   connection.commit()
in parallel with different keys; it could be a PersistentDict or a
PersistentList instead of btree. Should I write my own server that accepts
connections from all processes and serialize writing to the DB?

   Can I have a serial (autoincremented) counter? Should I? They are useful
in cases when I don't have an interesting distinguishing features in my
objects. For example, if I want to store a list of FTP servers I can
distinguish them by their URLs; hence I can use URLs as indices. But if I
want to store access_log elements - there are no such distinguishing
elements. Even the full tuple (time, client IP, URL) can occur many times
(many concurrent queries from a program like ApacheBenchmark; or a huge
network behind a NAT with a single external IP). Autoincremented counter
seems to be the best way to generate names (indices).
   The only counter I've found is .new_oid()...

   How can I know programmatically if packing is finished? Necessary for an
administration script that initiates pack, wait and backup .prepak file.

Oleg.
--
     Oleg Broytmann            http://phd.pp.ru/            phd@phd.pp.ru
           Programmers don't die, they just GOSUB without RETURN.