EMF Compare/Logical Model
What is a logical model?
We name "logical model" a set of physical resources that form a coherent business model. For example, we could say that a given Java class forms a coherent logical model only when it is linked with all of its imported classes.
In the case of EMF, we name a logical resource (or model) the EMF resource loaded in memory, as opposed to a physical resource (or file) that is merely the serialization of this model on disk. A given EMF model can reference a number of other models, and it will be incoherent, or even sometimes corrupted, if these other models are not loaded in memory. In EMF, a given model can be serialized as a single file, fragmented in multiple files on disk, or reference multiple files. The logical model is only coherent when the whole set of its physical files is accessible.
The Eclipse Team project (referred as "Team" in this document) provides an API named "model providers". This API allows implementers to define the semantics of what is a "logical model" in his case. In short, it allows us to link any number of physical resources to a given "starting" file.
Technically, this is done through an extension point that can be implemented by anyone and that will adapt a file (a workspace IResource) to a set of files (a Team ResourceTraversal). The model providers can be queried by anyone who calls actions on workspace files in order to determine if this action can be done against a single file or if it should be executed against a set of files.
Team only provides the API to define logical models. It is then the responsibility of its clients to properly query the model provider when calling an action. In the context of EMF Compare, we are interested in the "compare" actions. These actions are contributed by the repository providers (CVS, EGit, Subversive, Subclipse, Clearcase...). It is their responsibility, within the code of these actions, to query the model providers in order to determine if the selected files can be compared alone... or if they need to be compared along with a set of others.
- The CVS plugin does not consistently use it. For example, when using "compare with > latest from HEAD" on a file that is part of a logical model, it will 'see' the logical model and open the "synchronize" perspective instead of a compare editor : this is what's expected. However, when the user, from the synchronization view, asks to see the differences (right-click then select 'open in compare editor')... CVS will not query the logical model, comparing the file alone (see also bug 345415).
- The EGit, Subversive and Subclipse plugins never query the model providers for any comparison action.
In the case of EMF Compare, the above limitations were a show-stopper. EMF models cannot simply be reduced to single files as that would totally break their consistency when trying to compare them.
EMF Compare now uses its own implementation aiming at this same goal, that we named synchronization model. In short, instead of expecting the repository providers to query the model providers, EMF Compare will take the file (or set of files) that it is fed and expand it to the full logical model by querying the synchronization model.
This custom implementation also allows us to not only resolve the whole logical model of a given physical file, but also to determine which parts of this model should be included in the comparison scope. Indeed, under the semantics of EMF, if we can determine that two physical files are binary identical in their two distinct revisions, then we also know that the part of the logical model they represent are identical themselves. There would be no need to compare them. This allows to save both time and memory, killing two birds with one stone.
How it works
When the user selects a file A in his workspace and asks to compare it with a remote revision (for example, the "latest from HEAD"), the repository provider (CVS in this case) fetches the remote content of A in the selected revision (we'll call it A'). Once the remote revision fetched, the repository provider considers his work finished and feeds EMF Compare the two files (A and A') and lets it begin his own work. This is fine and dandy when the logical model of A only spans a single file ... but if A references another file B, then the comparison will not be coherent as EMF Compare operates on the logical model, not on its serialized form.
Thus, EMF Compare will not immediately launch the comparison of the files it is fed. Before any further work, it will query the synchronization model with these files in order to expand them to their whole logical model.
- In the case of a local file (A), it will load it as an EMF Resource and resolve all of its cross references. During the resolution process, we'll be able to track all of the "other" files required by the "starting point" A. In the example above, we'll end up with a set containing both A and B.
- In the case of a remote file (A'), it will try and load the streamed revision of the resource, then query the repository provider for the other files needed by A' in the correct revision. A revision is deemed "correct" if it precedes (or is equal to) the revision of A' that we have initially been fed. This will lead us to a set containing A' and B'.
Once the full logical model is resolved, EMF Compare can properly do the comparison work. Since we know of all fiels that compose our model, we'll also be able to safely merge differences without compromising the coherence of the whole.
However, even if we can launch EMF Compare at this point, the synchronization model also offers us one more possibility. EMF models can span thousands of physical files sometimes amounting to such a huge number of elements that it would not fit into the machine's memory if we tried to load it as a whole. The synchronization model knows how to retrieve the content of all of these files; it can also be used to minimize the set to only those which really did change. Indeed, we consider that files that are binary identical wield yield fragments of the logical model that will also be strictly identical. By removing these "identical" fragments, we can drastically limit the size of the logical model we'll load into memory, compacting it to a size that will fit before we launch the comparison on the remaining fragments.
The use of our own implementation effectively allows us to bypass the aforementioned limitations since we no longer rely on the repository providers. However, it has limitations of its own as far as model coherence is concerned.