Hi all!
Quoting Doug Way dway@riskmetrics.com: [SNIP]
Okay, this is getting into a difference between these two approaches which needs to be fleshed out some more (and which I'm somewhat familiar with, at least :-) ).
With Envy/Ginsu-style modules (which I assume are similar to what Roel &
Steph are talking about), modules can contain classes, and can also contain "Class Extensions" (a.k.a. loose methods). I would define Class
Extensions as methods which are additions to classes in other modules, but they are ONLY additions, NOT changes to existing methods (and not removals either).
DeltaModules, on the other hand, can contain method additions, but also
changes to existing methods and method removals too.
And in fact any other possible change too (just so that noone misunderstands that, I know you know this).
So, to generalize a bit, DeltaModules are more powerful, but Class Extensions are quite a bit simpler.
Let's think about a couple of useful properties of modules:
- Can they be cleanly unloaded?
- Are they behavior-preserving? In other words, if you load (and
activate) a module, will the rest of the system still run without breaking or behaving differently?
Regarding #1, 3.3a-style Modules and DeltaModules can be unloaded, and Envy/Ginsu-style modules can too, so no problems there. Changesets cannot be unloaded, which is their big limitation.
Regarding #2, this is one reason for the split between Modules and DeltaModules in 3.3a. If you load a Module (with no DeltaModules), the
rest of the system will be unaffected, so they are behavior-preserving.
DeltaModules are not behavior-preserving, but that's the point of them... they can be used if you do need to change something else in the
system.
Because DeltaModules are not behavior-preserving, they naturally have to
change version numbers of the Modules that they modify. So for example
if a WebBrowser Module has a DeltaModule which adds the method #asEncodedHtml or changes a method in class String, in the Module "Kernel 2.7", the Kernel module will need to be assigned a new version number such as 2.8. (I believe this is how DeltaModules are supposed to
work, correct me if I'm wrong.) This may have big implications for the
This is correct AFAIK.
rest of the modules in the system, which might expect to run on Kernel 2.7 but not know about 2.8.
Exactly.
But the interesting thing is, Envy/Ginsu-style modules w/Class Extensions *are* behavior-preserving, with a few caveats* (see bottom of message). This is because, in general, if you add a new method to an existing class, none of the rest of the system would ever send that new method, so the behavior of the rest of the system is unchanged. So, reusing the example above, if a WebBrowser module has a Class Extension which adds the method #asEncodedHtml (but does not change any method) in class String, in the module "Kernel 2.7", the Kernel module will still remain 2.7. (This is exactly what Envy does. I don't remember now if Ginsu supports version numbers.)
Yes, a very crucial observation. Just a note though: You listed a few caveats but there is one more - the new behaviour may of course affect the state of the instance so that other "modules" are affected. But that is probably more of a "bug" than a "source code conflict". Whatever.
The other question is then, if we're using Class Extensions, is it good enough to only be able to *add* methods to classes in other modules, not modify/remove them (as DeltaModules can do)?
Well, I guess that's a big question. :-) I would say that most of the time, yes, it's good enough. A module does not often need to change/remove methods in other modules, unless we're directly fixing a bug in that other module. Adding methods is much more useful, and can be specific to the "outer" module (such as the example WebBrowser module adding #asEncodedHtml to String). People who've used Envy know that it can be very handy... although you don't want to abuse it to an extreme.
With Envy/Ginsu, if your module *really* needs to change/remove a method in Kernel 2.7, you make your own private version of Kernel 2.8 with the change, and/or you lobby the maintainer of Kernel to make the change. But this should be relatively rare aside from bug fixes.
On the other hand, the ability of DeltaModules to handle method changes/removals could be quite nice. And they seem a bit more rigorous, no "caveats" as with Class Extensions below. And namespaces are already built-in. But there is extra complexity cost... For example, DeltaModules have to worry about activation/deactivation, with Class Extensions it is irrelevant. The Class Extensions are always active when the module is active, since they don't affect anything else.
One last thing: As far as splitting up the current Squeak image into modules goes, my hunch is that it would be easier to do with Envy/Ginsu-style modules. This is because you would not need the DeltaModule capability of specifying changed/removed methods when divvying things up, only the ability to specify added methods, which the lighter-weight Class Extensions can also handle without version number worries. For example, String>>asMorph could be quickly dumped into some Morphic module.
There might also be another problem lurking with DMs - at least when trying to use them to chop up the image. Since a DM branches off the affected base Module then it would seem to me that we will end up with tons of "branches" all over the place. Module C has a DM which branches off Kernel to 2.8 and Module D has another DM that branches off Kernel to 2.9 - oops...
(Avi, are the DVS "logical modules" similar to Envy/Ginsu with Class Extensions?)
Yes, that is my sense of DVS. But using category coventions and thus staying "compatible" with good old 3.2. But hey, I don't think DVS has any form of conflict resolution capabilites or anything like that.
Am I making sense here? This message is getting too long, sorry. Let me know if I've characterized one of the approaches incorrectly. And I know this is only one aspect of the differences between these two module systems. Anyway, I've often found that I learn more by see two things compared against each other, than by seeing them described separately.
You are making sense, and it was an interesting angle. Personally I think I have understood Henrik's Modules but I have always been wondering what will happen when we reach the real interesting area - conflict resolution...
The following are a few unstructured thoughts:
Currently I think that producing Envy/Ginsu style packages (for example using DVS) consisting of standalone classes and loose method extensions will work quite fine for many of us (as you also noted) - especially if you build stuff "for" 3.2, compared to trying to detangle 3.2 itself...
But I would like to have these "loose methods" contained within a new smarter kind of ChangeSet that I will call a "Patch" here below. I will try to explain why and what such a beast would be.
- A Patch would be similar to a DM in that it is not a recording of a sequence of changes - it is instead a condensed set of changes, much like a patch-file in Unix.
- On the other hand a Patch would not be a Module (as DMs are) or anything like that in itself. Instead a Patch is an "intermediate object" like DVS and other tools could produce when they want to represent "loose method additions" or "loose method removals" or "class instvar additions" etc. Patches would be able to represent themselves in a simple format that can be used inside fileouts and of course they would be instantiated when filing in such fileouts.
So the Patch would not live in the image like classes do but tools could produce them as an intermediate object when filing in or out. And when installing them it may be a good idea to keep them around.
Ok. A Patch would not try to replace ChangeSets in any way. A ChangeSet is a "recorded session" in some sense and thus IMHO more of a development tool than a packaging/deployment tool. I think we should keep using them for source code management but perhaps we should try to avoid using them for deployment/packaging and instead rely on .st files containing full classes and Patches.
Here are a few crucial points... A Patch object should:
1. Keep enough state about the changes so that it can be used to detect conflicts against other Patch objects and packages. For example: One Patch adding a loose method x in class A would not conflict with another Patch adding method y in class A - they would both be additions. A Patch changing a method x in class A (belonging to package AA) would on the other hand potentially conflict with a package BB that depends on package AA. But perhaps we could even add classifications of "method changes" - one being "no behaviour change" (changing comments, formatting, optimization etc) and then it could be resolved into a "theoretical conflict" (likely not a conflict but could be if the method change is actually changing behaviour nevertheless) instead of a "conflict". So in short - we can arm these Patch objects with a lot of "smarts" for detecting conflicts and also help us resolving them.
2. Keep enough state for deinstallation. This part is (as I have argued in other posts) not truly critical but if we try to implement these little fellows we could at least give it a stab. It would be nice.
Both of the above features would need the Patch objects to stay around after filing in, not necessarily in the image though - we could log them in a file or something.
So in short, the proposal of a lightweight module system could be like this:
1. We implement the Patch object as described above. 2. Avi :-) incorporates Patch objects into DVS. 3. We start using .st files with Patches inside them for packaging/deployment. One way would be by using Envy/Ginsu style packages with DVS. 4. We add dependencies to SqueakMap between packages. We would also need to add versions of packages of course because currently SqueakMap only knows about the latest release).
In Henrik's Modules dependencies were declared "per Module". Perhaps that level is too fine grained? If we add dependencies in SM it would be "per package" which is a larger and perhaps more appropriate level.
All this would mean that we can continue using ChangeSets if we like. Using them for deployment would lead to "uninstallable" packages with much less capability of conflict resolution.
Then when we try to install a package using SM it holds information on what packages we already have installed. Packages using .cs or "interactive" .st-files can't really participate in the conflict detection - hopefully those kind of packages will "phase out". But the "new" .st format with Patches complemented with dependency information in SM should be both uninstallable and especially conflict detection between them should work nicely.
In short: The new Patch object is IMHO the key for successful packages/modules. And the simplest Patch is a "loose method addition" which will give us Envy/Ginsu style modules. But then nothing stops us from evolving the Patch concept with more smarts.
Wow, what a long post this became...
regards, Göran
Göran Hultgren, goran.hultgren@bluefish.se GSM: +46 70 3933950, http://www.bluefish.se "Department of Redundancy department." -- ThinkGeek