[Jeremy Hylton]
> IIRC, the old implementation of savepoints kept a copy of the index at
> the time the savepoint was taken so that you could rollback to it
> multiple times.  I don't think there's any way to avoid such a copy.

Right, and the current implementation did that too.  The "surprise" was that
it wasn't enough.  Sketch:

1. Modify object 0.
2. Make savepoint 1.
       It makes a copy of the current index, say {0: 0}, and remembers
       the TmpStore size, say 100.
3. Fiddle around.
4. Rollback to savepoint 1.
       This sets the TmpStore index to the saved {0: 0}, and truncatss
       TmpStore to size 100.  So far so good -- or so it seems.
5. Modify object 0 again, and make savepoint 2.
       This changes the TmpStore index to {0: 100}, makes of a copy of
       {0: 100} in savepoint 2, and increases TmpStore size to 200.  This
       just did something horribly wrong too, although it's subtle.
6. Rollback to savepoint 1 again.
       Because a copy of savepoint 1's index wasn't _also_ made in
       step #4, the index savepoint 1 is holding onto mutated to
       {0: 100} during step #5 (object sharing).  This (#6) step
       leaves TmpStore with (the mutated) index {0: 100} and size 100.
7. Reference object 0.
       Oops.  The index tells us to seek to pos 100, but TmpStore has
       been truncated to 100.  We get a low-level exception from
       struct.unpack() about not enough bytes to unpack the data record

You can guess that I saw that happening <wink>.  Step #4 also needs to copy
the index (from the savepoint to TmpStore) instead of sharing a reference,
although this wasn't needed so long as a savepoint could be "used" at most
once (then mutating the savepoint's index after a rollback had no ill
effect, as the savepoint's index could never be referenced again).

For more information about ZODB, see the ZODB Wiki:

ZODB-Dev mailing list  -  ZODB-Dev@zope.org

Reply via email to