>should there be an undo stack per dataset, or one undo stack for
>the whole application?
>
>I think the latter.

My text editor has one undo stack per file, and this seems very natural to me.  Can you explain why you think that one global undo stack is more natural?

I expect that each data set will have its own top-level window.  Is that also your expectation?  In such a case, it seems to me that selecting "Undo" from one window and having a different data set change would be surprising.

>As I recall, what spss does (did) was to allow multiple levels of
>undo, until an EXECUTE happens, at which point the stack is
>emptied and one cannot undo anything.

We should be able to match that, at least, although it will take some work.

Most procedures don't modify the data set, so it would also be possible to extend the undo past those procedure.

>I can't think of any general and efficient way to undo the effect
>of an EXECUTE command.

General and efficient?  Me neither.  But it would be pretty easy to undo it in a general way, by keeping around the old casereader.  The need to preserve other parts of the dataset (e.g. dictionary) make it slightly nontrivial, but still it wouldn't be that hard.

The cost, at most, would be doubling the memory and disk requirements (for a single level of undo).

Do we run EXECUTE after every operation?  We wouldn't necessarily have to do that; for example, most transformations can be simulated on a per-case basis.  But that would not be a simple change (even though it would be convenient for users).