`exp run`: data gets re-imported on every call

[daavoo]

Bug Report

Description

When a pipeline uses an imported data file (with dvc import), the data gets cloned(?) and hashed every time dvc exp run is called.

Reproduce

1. dvc import git@github.com:iterative/dataset-registry.git use-cases/cats-dogs
2. dvc stage add -n foo -d cats-dogs echo foo
3. dvc exp run

Expected

When using dvc repro the imported data doesn't get re-hashed. I would expect dvc exp run to behave the same.

Environment information

Output of dvc doctor:

$ dvc doctor
DVC version: 2.6.3 (pip)
---------------------------------
Platform: Python 3.9.6 on macOS-10.16-x86_64-i386-64bit
Supports:
        gdrive (pydrive2 = 1.9.1),
        http (requests = 2.26.0),
        https (requests = 2.26.0)
Cache types: reflink, hardlink, symlink
Cache directory: apfs on /dev/disk1s1s1
Caches: local
Remotes: None
Workspace directory: apfs on /dev/disk1s1s1
Repo: dvc, git

Additional Information (if any):

$ dvc repro -v
2021-08-25 11:11:55,186 DEBUG: Computed stage: 'cats-dogs.dvc' md5: '5a135b297ee3c96465ce4b320f44fb8b'
'cats-dogs.dvc' didn't change, skipping
Stage 'foo' didn't change, skipping
Data and pipelines are up to date.

$ dvc exp run -v
2021-08-25 11:12:15,672 DEBUG: Detaching HEAD at 'HEAD'               
2021-08-25 11:12:15,690 DEBUG: Stashing workspace
2021-08-25 11:12:15,700 DEBUG: No changes to stash
2021-08-25 11:12:15,749 DEBUG: Creating external repo git@github.com:iterative/dataset-registry.git@ca140591a21c6d75a7057d1e2eb3f51d3115c5f5
2021-08-25 11:12:15,749 DEBUG: erepo: git clone 'git@github.com:iterative/dataset-registry.git' to a temporary dir
Computing file/dir hashes (only done once)  
. . .                                                                         

[pmrowla]

Issue occurs because we force dvc commit to preserve any initial changes to DVC data in the experiment. The problem is that the dvc commit implementation uses stage.changed_entries() to determine whether or not it needs to call stage.save()

dvc/dvc/repo/commit.py

Lines 57 to 65 in ed2ccaf

	for stage_info in stages_info:
	stage = stage_info.stage
	changes = stage.changed_entries()
	if any(changes):
	prompt_to_commit(stage, changes, force=force)
	stage.save(allow_missing=allow_missing)
	stage.commit(
	filter_info=stage_info.filter_info, allow_missing=allow_missing
	)

stage.changed_entries() returns workspace_status() which is fine for everything except imports. In the import case, workspace_status actually reports whether or not an upstream update is available, which requires this additional erepo clone to check the upstream, and is not particularly useful in the dvc commit case, except maybe as a warning to the user that they may have forgotten to do a dvc update.

In the even that we call commit with force=True we should definitely be skipping this status check for imports, and I'm not entirely sure it's still useful even when force=False