checkpoints: completely remove checkpoint outs on exp run --reset (#5586)

* checkpoints: remove checkpoint outs on --reset

* exp run: no longer prune/reset lockfiles

* update tests

* make --queue imply --reset unless --rev is provided

dvc/command/experiments.py

@@ -514,6 +514,11 @@ class CmdExperimentsRun(CmdRepro):
     def run(self):
         from dvc.command.metrics import _show_metrics
 
+        if self.args.reset and self.args.checkpoint_resume:
+            raise InvalidArgumentError(
+                "--reset and --rev are mutually exclusive."
+            )
+
         if self.args.reset:
             logger.info("Any existing checkpoints will be reset and re-run.")
 

dvc/output/base.py

@@ -408,6 +408,7 @@ class BaseOutput:
         relink=False,
         filter_info=None,
         allow_missing=False,
+        checkpoint_reset=False,
         **kwargs,
     ):
         if not self.use_cache:
@@ -422,6 +423,11 @@ class BaseOutput:
             # backward compatibility
             return None
 
+        if self.checkpoint and checkpoint_reset:
+            if self.exists:
+                self.remove()
+            return None
+
         added = not self.exists
 
         try:

dvc/repo/experiments/__init__.py

@@ -182,8 +182,6 @@ class Experiments:
                         )
                         self.scm.reset()
 
-                    self._prune_lockfiles()
-
                     # update experiment params from command line
                     if params:
                         self._update_params(params)
@@ -244,34 +242,6 @@ class Experiments:
 
         return stash_rev
 
-    def _prune_lockfiles(self):
-        from dvc.dvcfile import is_lock_file
-
-        # NOTE: dirty DVC lock files must be restored to index state to
-        # avoid checking out incorrect persist or checkpoint outs
-        fs = self.scm.get_fs("HEAD")
-        lock_files = [
-            str(fname)
-            for fname in fs.walk_files(self.scm.root_dir)
-            if is_lock_file(fname)
-        ]
-        if lock_files:
-
-            self.scm.reset(paths=lock_files)
-            self.scm.checkout_index(paths=lock_files, force=True)
-
-    def _prune_untracked_lockfiles(self):
-        from dvc.dvcfile import is_lock_file
-        from dvc.utils.fs import remove
-
-        untracked = [
-            fname
-            for fname in self.scm.untracked_files()
-            if is_lock_file(fname)
-        ]
-        for fname in untracked:
-            remove(fname)
-
     def _stash_msg(
         self,
         rev: str,
@@ -345,9 +315,17 @@ class Experiments:
         queue: bool = False,
         tmp_dir: bool = False,
         checkpoint_resume: Optional[str] = None,
+        reset: bool = False,
         **kwargs,
     ):
         """Reproduce and checkout a single experiment."""
+        if queue and not checkpoint_resume:
+            reset = True
+
+        if reset:
+            self.reset_checkpoints()
+            kwargs["force"] = True
+
         if not (queue or tmp_dir):
             staged, _, _ = self.scm.status()
             if staged:
@@ -370,7 +348,9 @@ class Experiments:
         else:
             checkpoint_resume = self._workspace_resume_rev()
 
-        stash_rev = self.new(checkpoint_resume=checkpoint_resume, **kwargs)
+        stash_rev = self.new(
+            checkpoint_resume=checkpoint_resume, reset=reset, **kwargs
+        )
         if queue:
             logger.info(
                 "Queued experiment '%s' for future execution.", stash_rev[:7],
@@ -709,7 +689,6 @@ class Experiments:
         # result in conflict between workspace params and stashed CLI params).
         self.scm.reset(hard=True)
         with self.scm.detach_head(entry.rev):
-            self._prune_untracked_lockfiles()
             rev = self.stash.pop()
             self.scm.set_ref(EXEC_BASELINE, entry.baseline_rev)
             if entry.branch:

dvc/repo/experiments/executor/base.py

@@ -259,17 +259,19 @@ class BaseExecutor(ABC):
                 "Executor repro with force = '%s'", str(repro_force)
             )
 
-            # NOTE: for checkpoint experiments we handle persist outs slightly
-            # differently than normal:
+            # NOTE: checkpoint outs are handled as a special type of persist
+            # out:
             #
             # - checkpoint out may not yet exist if this is the first time this
             #   experiment has been run, this is not an error condition for
             #   experiments
-            # - at the start of a repro run, we need to remove the persist out
-            #   and restore it to its last known (committed) state (which may
-            #   be removed/does not yet exist) so that our executor workspace
-            #   is not polluted with the (persistent) out from an unrelated
-            #   experiment run
+            # - if experiment was run with --reset, the checkpoint out will be
+            #   removed at the start of the experiment (regardless of any
+            #   dvc.lock entry for the checkpoint out)
+            # - if run without --reset, the checkpoint out will be checked out
+            #   using any hash present in dvc.lock (or removed if no entry
+            #   exists in dvc.lock)
+            checkpoint_reset = kwargs.pop("reset", False)
             dvc_checkout(
                 dvc,
                 targets=targets,
@@ -277,6 +279,7 @@ class BaseExecutor(ABC):
                 force=True,
                 quiet=True,
                 allow_missing=True,
+                checkpoint_reset=checkpoint_reset,
             )
 
             checkpoint_func = partial(

dvc/repo/experiments/run.py

@@ -15,7 +15,6 @@ def run(
     run_all: bool = False,
     jobs: int = 1,
     tmp_dir: bool = False,
-    reset: bool = False,
     **kwargs,
 ) -> dict:
     """Reproduce the specified targets as an experiment.
@@ -25,10 +24,6 @@ def run(
     Returns a dict mapping new experiment SHAs to the results
     of `repro` for that experiment.
     """
-    if reset:
-        repo.experiments.reset_checkpoints()
-        kwargs["force"] = True
-
     if run_all:
         return repo.experiments.reproduce_queued(jobs=jobs)
 

tests/func/experiments/test_checkpoints.py

@@ -88,10 +88,6 @@ def test_reset_checkpoint(
         checkpoint_stage.addressing, name="foo", tmp_dir=not workspace,
     )
 
-    if workspace:
-        scm.reset(hard=True)
-        scm.gitpython.repo.git.clean(force=True)
-
     results = dvc.experiments.run(
         checkpoint_stage.addressing,
         params=["foo=2"],

tests/func/experiments/test_experiments.py

@@ -418,28 +418,6 @@ def test_untracked(tmp_dir, scm, dvc, caplog, workspace):
         assert fobj.read().strip() == "foo: 2"
 
 
-@pytest.mark.parametrize("workspace", [True, False])
-def test_dirty_lockfile(tmp_dir, scm, dvc, exp_stage, workspace):
-    from dvc.dvcfile import LockfileCorruptedError
-
-    tmp_dir.gen("dvc.lock", "foo")
-
-    with pytest.raises(LockfileCorruptedError):
-        dvc.reproduce(exp_stage.addressing)
-
-    results = dvc.experiments.run(
-        exp_stage.addressing, params=["foo=2"], tmp_dir=not workspace
-    )
-    exp = first(results)
-
-    fs = scm.get_fs(exp)
-    with fs.open(tmp_dir / "metrics.yaml") as fobj:
-        assert fobj.read().strip() == "foo: 2"
-
-    if not workspace:
-        assert (tmp_dir / "dvc.lock").read_text() == "foo"
-
-
 def test_packed_args_exists(tmp_dir, scm, dvc, exp_stage, caplog):
     from dvc.repo.experiments.executor.base import BaseExecutor