diff: consider including modified git-tracked files

[shcheklein]

If the purpose of the dvc diff is to take a glance at what has changed in my iteration vs HEAD, or compare two experiments, it would be really convenient to include output of git diff --name-only in the output.

E.g. when I run dvc diff baseline-experiment bigrams-experiment, I get output like:

Modified:
    auc.metric
    data/features/
    data/features/test.pkl
    data/features/train.pkl
    model.pkl

files summary: 0 added, 0 deleted, 4 modified

it's fine, but can be misleading or not very informative. In this specific case there is also a change in the script train.py itself which is an essential part of the pipeline. I think, it would make to see at the name here.

Any though @iterative/engineering ?

UPDATE: Giving a second thought to this, I see that it's not even about DVC-tracked (cached) files vs Git-tracked files but DVC outputs vs outputs+dependencies vs all files (DVC and Git).

So, three options:

1. Only changes to DVC "outputs" (e.g. including the Git-tracked metric file like in the example) above;
2. Changes to DVC outputs and dependencies (that we take from DVC-files);
3. All Git and DVC files.

[casperdcl]

I'd vote 2.

3 if specified via a --git flag, embellishing git-tracked files with an asterisk/colour

[jorgeorpinel]

I'd vote for 2. as well, but maybe also 3. if dvc install has been run ("full Git integration mode").

---

Other possibly unnecessary comments

the purpose of the dvc diff is to take a glance at what has changed in my iteration vs HEAD

This seems like the tricky case: the default behavior, where dvc diff (without args) compares current workspace (working tree, including "local changes that you have made but not yet committed") -minus- HEAD.

it's not even about DVC-tracked (cached) files vs Git-tracked files but DVC outputs vs outputs+dependencies vs all files (DVC and Git).

I think it's actually about [ DVC-tracked but (Git) uncommitted changes in current workspace ] -minus- HEAD (current behavior)
vs. any of the other options (2. and 3.)

This is in contrast to dvc status which checks workspace changes not yet tracked by DVC.
Related comment: treeverse/dvc.org#953 (comment)

[dmpetrov]

Now we have:
0. Diff only for data files and dirs changes.

All the options (probably exclude (1)) might be very convenient in different cases. The major question is - what should be used as a default. If we include Git file changes in the default output then we need to provide clear marks in the output what is data file/dir and what is not. If the default is (0) then it might be ok not to supply this mark (in JSON-format it might be still valuable).

[dmpetrov]

I'd vote for (2) if we come up with a good formating for Git/Data files. Otherwise - (0).
In any case, it worth to keep all (probably except (1) ) through options.

[shcheklein]

@dmpetrov 0 - meaning no changes to metric files (no matter cached or not)?

[dmpetrov]

@dmpetrov 0 - meaning no changes to metric files (no matter cached or not)?

It supposed to work for changed data files including metrics, right?

[shcheklein]

@dmpetrov I'm just trying to understand what is the difference between 0 and 1 then.

[dmpetrov]

0 is the current behavior

[shcheklein]

@dmpetrov yep, that's what I meant by 1 :) may be I missed something important about the current behavior though, but I didn't mean to change it - 1 is the current one, at least the way I understood it initially.