[Dev] Undo and commit()

[Reid Ellis]

On Tue Jan 3 2006, at 14:22, John Anderson wrote:
> Alec Flett wrote:
>> John Anderson wrote:
>>> I was hoping, someday, to have the commit point be the unit of  
>>> undo, e.g. if you did a delete, you'd commit after the operation.  
>>> Of course this requires the ability to roll back changes to  
>>> previous commit points, which doesn't yet exist.
>>
>> Personally, I think that trying to tie Undo transactions to  
>> repository transactions is going to land us in a heap of trouble.  
>> I'm all for thinking of Undo-able actions as a series of  
>> repository changes, but there is a disconnect between what the  
>> user probably considers "undoable" events and what the repository  
>> considers uncommittable.

Most undo/redo mechanisms I've dealt with:

	(a) are not tied too tightly to the data representation (using  
database rollback as the only method of undo)

	(b) do not act on UI changes (e.g. changing from week to day view)

I would argue that undo/redo should be achieved by parcelling up the  
code into "actions", similar to what Alec described, that know how to  
change the state of the model in both directions, thus being undoable  
and redoable. Based on what I've seen about how the repository is  
tied to both UI actions and external actions (like new mail), I do  
not think it would be a good idea to tie the undo architecture to the  
repository. Rather, they would make additional repository changes in  
order to achieve their required state change.

Perhaps as an optimization, they could check for external changes,  
and if there are none, then use the repository's rollback mechanism,  
but even that seems difficult, based on what Andi said about not  
being able to discard changes (or is that something that is coming  
soon?). Regardless, it goes out the window if something else happens  
like getting mail in the middle of doing undo/redo.

Is the "rerender" mechanism actually just marking items as "dirty",  
or is it forcing them to draw?

Reid

>>
>> Here are a few scenarios that I think would cause us trouble if we  
>> were to tie undo points directly to calls to commit()/uncommit().  
>> I'm using the term "undoable change" to be a user-comprehensible  
>> change, such as deleting an event or changing data in the detail  
>> view - one that the user would consider "undoable" by selecting  
>> Edit->Undo. I'm using the term "uncommit" to refer to playing back  
>> a series of changes in reverse order - this may or may not mean  
>> simply rolling back, as you'll see below.
>>
>> Scenario 1: an undoable change may actually span multiple commits
>> A user makes an undoable change that causes commit() to be called  
>> three times just because of the codepath it follows happens to  
>> have an extra commit() or two.. (maybe it goes through sharing and  
>> that causes some repository view manipulation) Now any sort of  
>> "Undo" command would have to be run three times to uncommit that  
>> single user action
> In past systems that used a mechanism like I'm proposing, you can  
> avoid this problem quite easily. Organize your code into operations  
> that do work, but don't commit. A command calls one or more  
> routines that the work then the command does a commit at the end  
> (Actually, to be more precise, each time a command was begun it  
> commited the last commands changes)
>>
>> Scenario 2: A user manipulates the UI after making an undoable change
>> A user makes an undoable change, and then clicks on another  
>> collection. Thanks to CPIA, changes in the UI are also changes in  
>> the repository. This means that "Undo" operation might actually  
>> just cause the user to switch back to their preview collection. It  
>> would take a 2nd Undo operation to actually "undo" the original  
>> change.
> I'm not sure I understand this, but if you mean that undo moves the  
> user interface back to exactly what it was when you did the last  
> commit, yes that's the way it works and this is what you expect --  
> undo should get you back to exactly where you were, view included.
>>
>> Scenario 3: Some changes might occur in different repository views
>> A user makes an undoable change, and that change may cause changes  
>> in multiple views - i.e. sharing. What exactly do we "undo" when  
>> we uncommit?
> If I understand this, another variation of it is the following:  
> Suppose while you're doing some operation, new mail arrives, when  
> you undo you don't want to undo the getting of the new mail. This  
> is a real problem. One possible solution is to save information  
> about changes to the repository from other threads and when you  
> undo (roll back) replay (merge) those changes into the repository.
>>
>> Personally, I think we need to be very explicit about undoable  
>> changes. I think we need to bracket such changes with some sort of  
>> begin/end undoable transaction mechanism. We can still exploit all  
>> the benefits of the repository though - for instance:
> Currently, there is no obvious rule about when to do a commit, so  
> it ends up being a little arbitrary. This would eliminate the any  
> confusion about when to do commit.
>> 1) an undoable transaction might just be a pair of repository  
>> version numbers - and we just play back all the changes backwards  
>> from the newest version to the oldest. (i..e even if we're at  
>> version 103, the previous undoable change might be from revision  
>> 100 to 101, so we'd just play the changes backwards between 101  
>> and 100)
> Yes, I agree something like this makes sense
>> 2) notifications would fire when we play back the changes, so the  
>> UI would (hopefully) stay up to date
> To keep the UI up to date, all we need to do is unrender the UI,  
> roll back the repository, then rerender it (which also calls  
> synchronizeWiget)
>> 3) perhaps we could even look ahead at the changes to be played  
>> back, to see if an undo operation is even valid. (i.e. if it might  
>> cause conflicts or something)
>> 4) Redo support would be easy - even multiple redo support like in  
>> Word/etc.
>>
>> An explicit Undo mechanism gives us another advantage: we could  
>> assign user-visible names to the undo actions.. and then we could  
>> display that user-visible name in the edit menu, so it might say  
>> "Undo Cut" instead of just "Undo"
> In another system I used based on a begin/end mechanism like  
> commit, commit took an argument which was the name of the command,  
> e.g. "Cut", which was used to keep the menu up to date.
>>
>> Here's an example of this system in action:
>>
>> UndoManager.BeginTransaction(_(u"Cut"))
>> self.CutSomething()
>> UndoManger.EndTransaction()
> If your BeginTransaction happened to call commit, I think we could  
> eliminate explicit commits and our proposals would be identical.
>>
>> Maybe there's even a more pythonic way to handle this:
>>
>> UndoManager.DoTransaction(_(u"Cut"), self.CutSomething)