> A related point:  Did we decide on write-through or explicit-commit
> semantics for attributes?
>
> There was discussion that explicit-commit semantics were
> needed, rather
> than the write-through semantics in Dietmar's original
> implementation (if I
> read the code correctly:-)  But was it ever resolved for sure
> and if so
> what is the rationale?
>
> Is the rationale just efficiency over slow connections, or is
> there also a
> fear that write-through of individual attributes could create an
> inconsistent file state (and thus get tossed by the OS)?
>

I'm definitely afraid of them getting tossed. The only way I would support write-through is if there's a way to set attributes en mass.  This would probably include the ability to:

1. copy attributes from a file object to a temporary that does not refer to a real file
2. adjust the attributes on the temporary
3. set the attributes of another file from the temporary "collection" of attributes.