[MERGE] filtered-deltas - basis for log DIR

[John Arbash Meinel]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1


...

>>> It seems strange for an optimization to be doing a copy.
>> Non-chk inventories are mutable so I'm doing the copying to
>> be safe.
> 
> You mean, because someone might mutate the inventory of the returned
> RevisionTree?  That's definitely not acceptable behaviour for using a
> RevisionTree.
> 
> Or do you mean someone using filter?  That could be addressed by
> rejigging filter so it used the current Inventory instead of returning a
> new one.
> 

We have some code paths that do something like:

tree = repo.revision_tree()
inv = tree.inventory
...

ie = inv[file_id]
ie.foo = bar

save_new_inventory(inv)

The one I know the most about is the base
Repository.add_inventory_by_delta():

basis_tree = self.revision_tree(basis_revision_id)
basis_tree.lock_read()
try:
    # Note that this mutates the inventory of basis_tree, which not all
    # inventory implementations may support: A better idiom would be to
    # return a new inventory, but as there is no revision tree cache in
    # repository this is safe for now - RBC 20081013
    basis_inv = basis_tree.inventory
    basis_inv.apply_delta(delta)
    basis_inv.revision_id = new_revision_id
    return (self.add_inventory(new_revision_id, basis_inv, parents),
            basis_inv)
finally:
    basis_tree.unlock()


At least it has a nice big warning about it. We also have some similar
code in the test suite. Which is why returning exact inventory entries
from _unpack_entry() is currently disabled, even though it save ~25% of
the time for converting from an XML repo into a CHK one. (Note that I
enable it when converting, because I know that code path is safe...)

John
=:->

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (Cygwin)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iEYEARECAAYFAknDpc8ACgkQJdeBCYSNAAMHGgCgyTzFg7LSDjyy6L71BipsgvxA
LscAn29vpHoBOqm+eZbEFTXgvG4I14hD
=+NoF
-----END PGP SIGNATURE-----