[Tune] TrialRunner checkpointing shouldn't fail if ray.data.Dataset…

… w/o lineage captured in trial config (ray-project#33565)

* Fix dataset serialization if no lineage

Signed-off-by: Justin Yu <[email protected]>

* add a unit test

Signed-off-by: Justin Yu <[email protected]>

* Add data dependencies for the test

Signed-off-by: Justin Yu <[email protected]>

---------

Signed-off-by: Justin Yu <[email protected]>
Signed-off-by: Jack He <[email protected]>

.buildkite/pipeline.ml.yml

@@ -233,7 +233,7 @@
   instance_size: medium
   commands:
     - cleanup() { if [ "${BUILDKITE_PULL_REQUEST}" = "false" ]; then ./ci/build/upload_build_info.sh; fi }; trap cleanup EXIT
-    - TUNE_TESTING=1 ./ci/env/install-dependencies.sh
+    - TUNE_TESTING=1 DATA_PROCESSING_TESTING=1 ./ci/env/install-dependencies.sh
     - ./ci/env/env_info.sh
     - bazel test --config=ci $(./ci/run/bazel_export_options) --build_tests_only
       --test_tag_filters=medium_instance,-py37,-soft_imports,-gpu_only,-rllib,-multinode

python/ray/tune/execution/experiment_state.py

@@ -204,6 +204,11 @@ def checkpoint(
         # Checkpoint
         checkpoint_time_start = time.monotonic()
 
+        # NOTE: This context manager is for Ray Datasets captured in a trial config.
+        # This is the case when *tuning over datasets*.
+        # If the datasets have already been full executed, then serializing
+        # block refs means that this checkpoint is not usable in a new Ray cluster.
+        # This context will serialize the dataset execution plan instead, if available.
         with out_of_band_serialize_dataset():
             save_fn()
 

python/ray/tune/impl/out_of_band_serialize_dataset.py

@@ -17,7 +17,7 @@ def _reduce(ds: ray.data.Dataset):
         if "serialize_lineage" in tb:
             _already_in_out_of_band_serialization = True
             break
-    if not _already_in_out_of_band_serialization:
+    if not _already_in_out_of_band_serialization and ds.has_serializable_lineage():
         return _deserialize_and_fully_execute_if_needed, (ds.serialize_lineage(),)
     else:
         return ds.__reduce__()

python/ray/tune/tests/test_trial_runner_3.py

@@ -2,6 +2,7 @@
 from collections import Counter
 import logging
 import os
+import pandas as pd
 import pickle
 import shutil
 import sys
@@ -14,6 +15,7 @@
 import ray
 from ray.air import CheckpointConfig
 from ray.air.execution import PlacementGroupResourceManager, FixedResourceManager
+from ray.exceptions import OwnerDiedError
 from ray.rllib import _register_all
 from ray.rllib.algorithms.callbacks import DefaultCallbacks
 
@@ -1155,6 +1157,74 @@ def testPeriodicCloudCheckpointSyncTimeout(self):
             assert any("did not finish running within the timeout" in x for x in buffer)
             assert syncer.sync_up_counter == 2
 
+    def testExperimentCheckpointWithDatasets(self):
+        """Test trial runner checkpointing where trials contain Ray Datasets.
+        When possible, a dataset plan should be saved (for read_* APIs).
+        See `Dataset.serialize_lineage` for more information.
+
+        If a dataset cannot be serialized, an experiment checkpoint
+        should still be created. Users can pass in the dataset again by
+        re-specifying the `param_space`.
+        """
+        ray.init(num_cpus=2)
+        # Save some test data to load
+        data_filepath = os.path.join(self.tmpdir, "test.csv")
+        pd.DataFrame({"x": list(range(10))}).to_csv(data_filepath)
+
+        def create_trial_config():
+            return {
+                "datasets": {
+                    "with_lineage": ray.data.read_csv(data_filepath),
+                    "no_lineage": ray.data.from_items([{"x": i} for i in range(10)]),
+                }
+            }
+
+        resolvers = create_resolvers_map()
+        config_with_placeholders = inject_placeholders(create_trial_config(), resolvers)
+        trial = Trial(
+            "__fake",
+            experiment_path=self.tmpdir,
+            config=config_with_placeholders,
+        )
+        trial.init_local_path()
+        runner = TrialRunner(
+            experiment_path=self.tmpdir, placeholder_resolvers=resolvers
+        )
+        runner.add_trial(trial)
+        # Req: TrialRunner checkpointing shouldn't error
+        runner.checkpoint(force=True)
+
+        # Manually clear all block refs that may have been created
+        ray.shutdown()
+        ray.init(num_cpus=2)
+
+        new_runner = TrialRunner(experiment_path=self.tmpdir)
+        new_runner.resume()
+        [loaded_trial] = new_runner.get_trials()
+        loaded_datasets = loaded_trial.config["datasets"]
+
+        # Req: The deserialized dataset (w/ lineage) should be usable.
+        assert [el["x"] for el in loaded_datasets["with_lineage"].take()] == list(
+            range(10)
+        )
+        # Req: The deserialized dataset (w/o lineage) should NOT be usable.
+        with self.assertRaises(OwnerDiedError):
+            loaded_datasets["no_lineage"].take()
+
+        replaced_resolvers = create_resolvers_map()
+        inject_placeholders(create_trial_config(), replaced_resolvers)
+
+        respecified_config_runner = TrialRunner(
+            placeholder_resolvers=replaced_resolvers,
+            local_checkpoint_dir=self.tmpdir,
+        )
+        respecified_config_runner.resume()
+        [loaded_trial] = respecified_config_runner.get_trials()
+        ray_ds_no_lineage = loaded_trial.config["datasets"]["no_lineage"]
+
+        # Req: The dataset (w/o lineage) can be re-specified and is usable after.
+        assert [el["x"] for el in ray_ds_no_lineage.take()] == list(range(10))
+
 
 class FixedResourceTrialRunnerTest3(TrialRunnerTest3):
     def _resourceManager(self):