Auto dependency calculation --auto-dag

[dmpetrov]

Currently DVC implicitly derives dependencies based on command line args.
$ dvc run python mycode.py data/input.csv data/output.p
It goes like this:

1. If arg (let say data/output.p) looks like data file (in data dir) we assume it is an input parameter.
2. If a data file was changed we change our mind and say the arg is an output.

Also, we have a special args like --input or -i and --output or -o for explicit input\output definition.

Should we keep auto\implicit dependency calculation or we should hide this logic behind, let say --auto-dag or -a option?

Result will be:
$ dvc run --auto-dag python mycode.py data/input.csv data/output.p
or explicitly:
$ dvc run -i data/input.csv -o data/output.p python mycode.py data/input.csv data/output.p

[dmpetrov]

Actually, we use two step dependency calculation:

1. Command line based.
2. Repository state based (before and after command run).

Both of them make DVC logic quite fragile.

To make DVC more persistent we should probably turn off both of the options by default. New parameters --auto-args, --auto-git or combined --auto-dag will include one of these.

And best practice should be - specify your inputs and outputs explicitly.

[dmpetrov]

One more specific case #154

[efiop]

Add --no-auto-dag as a temporary solution.

[efiop]

No longer relevant, as we switched to explicit dependencies.