Fixed hyperopt trial syncing to remote filesystems for Ray 2.0 (#2617)

ludwig/hyperopt/execution.py

@@ -44,11 +44,11 @@
 if _ray_200:
     from ray.air import Checkpoint
     from ray.tune.search import SEARCH_ALG_IMPORT
-    from ray.tune.syncer import get_node_to_storage_syncer, SyncConfig
+
+    from ludwig.hyperopt.syncer import RemoteSyncer
 else:
     from ray.ml import Checkpoint
     from ray.tune.suggest import SEARCH_ALG_IMPORT
-    from ray.tune.syncer import get_cloud_sync_client
 
 
 logger = logging.getLogger(__name__)
@@ -778,12 +778,14 @@ def run_experiment_trial(config, local_hyperopt_dict, checkpoint_dir=None):
             )
 
         if has_remote_protocol(output_directory):
-            run_experiment_trial = tune.durable(run_experiment_trial)
-            self.sync_config = tune.SyncConfig(sync_to_driver=False, upload_dir=output_directory)
             if _ray_200:
-                self.sync_client = get_node_to_storage_syncer(SyncConfig(upload_dir=output_directory))
+                self.sync_client = RemoteSyncer()
+                self.sync_config = tune.SyncConfig(upload_dir=output_directory, syncer=self.sync_client)
             else:
-                self.sync_client = get_cloud_sync_client(output_directory)
+                raise ValueError(
+                    "Syncing to remote filesystems with hyperopt is not supported with ray<2.0, "
+                    "please upgrade to ray>=2.0"
+                )
             output_directory = None
         elif self.kubernetes_namespace:
             from ray.tune.integration.kubernetes import KubernetesSyncClient, NamespacedKubernetesSyncer

ludwig/hyperopt/syncer.py

@@ -0,0 +1,34 @@
+from typing import Any, Callable, Dict, List, Optional, Tuple
+
+from ray.tune.syncer import _BackgroundSyncer
+
+from ludwig.utils.data_utils import use_credentials
+from ludwig.utils.fs_utils import delete, download, upload
+
+
+class RemoteSyncer(_BackgroundSyncer):
+    def __init__(self, sync_period: float = 300.0, creds: Optional[Dict[str, Any]] = None):
+        super().__init__(sync_period=sync_period)
+        self.creds = creds
+
+    def _sync_up_command(self, local_path: str, uri: str, exclude: Optional[List] = None) -> Tuple[Callable, Dict]:
+        with use_credentials(self.creds):
+            return upload, dict(lpath=local_path, rpath=uri)
+
+    def _sync_down_command(self, uri: str, local_path: str) -> Tuple[Callable, Dict]:
+        with use_credentials(self.creds):
+            return download, dict(rpath=uri, lpath=local_path)
+
+    def _delete_command(self, uri: str) -> Tuple[Callable, Dict]:
+        with use_credentials(self.creds):
+            return delete, dict(url=uri, recursive=True)
+
+    def __reduce__(self):
+        """We need this custom serialization because we can't pickle thread.lock objects that are used by the
+        use_credentials context manager.
+
+        https://docs.ray.io/en/latest/ray-core/objects/serialization.html#customized-serialization
+        """
+        deserializer = RemoteSyncer
+        serialized_data = (self.sync_period, self.creds)
+        return deserializer, serialized_data

ludwig/utils/fs_utils.py

@@ -202,6 +202,16 @@ def delete(url, recursive=False):
     return fs.delete(path, recursive=recursive)
 
 
+def upload(lpath, rpath):
+    fs, path = get_fs_and_path(rpath)
+    pyarrow.fs.copy_files(lpath, path, destination_filesystem=pyarrow.fs.PyFileSystem(pyarrow.fs.FSSpecHandler(fs)))
+
+
+def download(rpath, lpath):
+    fs, path = get_fs_and_path(rpath)
+    pyarrow.fs.copy_files(path, lpath, source_filesystem=pyarrow.fs.PyFileSystem(pyarrow.fs.FSSpecHandler(fs)))
+
+
 def checksum(url):
     fs, path = get_fs_and_path(url)
     return fs.checksum(path)

tests/integration_tests/test_hyperopt.py

@@ -15,7 +15,8 @@
 import contextlib
 import json
 import os.path
-from typing import Any, Dict, Optional, Tuple, Union
+import uuid
+from typing import Any, Dict, Optional, Tuple
 
 import pytest
 import torch
@@ -39,23 +40,21 @@
 from ludwig.globals import HYPEROPT_STATISTICS_FILE_NAME
 from ludwig.hyperopt.results import HyperoptResults
 from ludwig.hyperopt.run import hyperopt, update_hyperopt_params_with_defaults
+from ludwig.utils import fs_utils
 from ludwig.utils.data_utils import load_json
 from ludwig.utils.defaults import merge_with_defaults
-from tests.integration_tests.utils import category_feature, generate_data, text_feature
+from tests.integration_tests.utils import category_feature, generate_data, private_param, remote_tmpdir, text_feature
 
-try:
-    import ray
+ray = pytest.importorskip("ray")
 
-    from ludwig.hyperopt.execution import get_build_hyperopt_executor
+from ludwig.hyperopt.execution import get_build_hyperopt_executor  # noqa
 
-    _ray113 = version.parse(ray.__version__) > version.parse("1.13")
+_ray200 = version.parse(ray.__version__) >= version.parse("2.0")
 
-except ImportError:
-    ray = None
-    _ray113 = None
+pytestmark = pytest.mark.distributed
 
 
-RANDOM_SEARCH_SIZE = 4
+RANDOM_SEARCH_SIZE = 2
 
 HYPEROPT_CONFIG = {
     "parameters": {
@@ -165,18 +164,6 @@ def _setup_ludwig_config_with_shared_params(dataset_fp: str) -> Tuple[Dict, Any]
     return config, rel_path, num_filters_search_space, embedding_size_search_space, reduce_input_search_space
 
 
-def _get_trial_parameter_value(parameter_key: str, trial_row: str) -> Union[str, None]:
-    """Returns the parameter value from the Ray trial row, which has slightly different column names depending on
-    the version of Ray. Returns None if the parameter key is not found.
-
-    TODO(#2176): There are different key name delimiters depending on Ray version. The delimiter in future versions of
-    Ray (> 1.13) will be '/' instead of '.' Simplify this as Ray is upgraded.
-    """
-    if _ray113:
-        return trial_row[f"config/{parameter_key}"]
-    return trial_row[f"config.{parameter_key}"]
-
-
 @contextlib.contextmanager
 def ray_start(num_cpus: Optional[int] = None, num_gpus: Optional[int] = None):
     res = ray.init(
@@ -198,7 +185,6 @@ def ray_cluster():
         yield
 
 
-@pytest.mark.distributed
 @pytest.mark.parametrize("search_alg", SEARCH_ALGS_FOR_TESTING)
 def test_hyperopt_search_alg(
     search_alg, csv_filename, tmpdir, ray_cluster, validate_output_feature=False, validation_metric=None
@@ -249,7 +235,6 @@ def test_hyperopt_search_alg(
         assert isinstance(path, str)
 
 
-@pytest.mark.distributed
 def test_hyperopt_executor_with_metric(csv_filename, tmpdir, ray_cluster):
     test_hyperopt_search_alg(
         "variant_generator",
@@ -261,7 +246,6 @@ def test_hyperopt_executor_with_metric(csv_filename, tmpdir, ray_cluster):
     )
 
 
-@pytest.mark.distributed
 @pytest.mark.parametrize("scheduler", SCHEDULERS_FOR_TESTING)
 def test_hyperopt_scheduler(
     scheduler, csv_filename, tmpdir, ray_cluster, validate_output_feature=False, validation_metric=None
@@ -316,7 +300,6 @@ def test_hyperopt_scheduler(
         assert isinstance(raytune_results, HyperoptResults)
 
 
-@pytest.mark.distributed
 @pytest.mark.parametrize("search_space", ["random", "grid"])
 def test_hyperopt_run_hyperopt(csv_filename, search_space, tmpdir, ray_cluster):
     input_features = [
@@ -370,14 +353,19 @@ def test_hyperopt_run_hyperopt(csv_filename, search_space, tmpdir, ray_cluster):
         "goal": "minimize",
         "output_feature": output_feature_name,
         "validation_metrics": "loss",
-        "executor": {TYPE: "ray", "num_samples": 1 if search_space == "grid" else RANDOM_SEARCH_SIZE},
+        "executor": {
+            TYPE: "ray",
+            "num_samples": 1 if search_space == "grid" else RANDOM_SEARCH_SIZE,
+            "max_concurrent_trials": 1,
+        },
         "search_alg": {TYPE: "variant_generator"},
     }
 
     # add hyperopt parameter space to the config
     config[HYPEROPT] = hyperopt_configs
 
-    hyperopt_results = hyperopt(config, dataset=rel_path, output_directory=tmpdir, experiment_name="test_hyperopt")
+    experiment_name = f"test_hyperopt_{uuid.uuid4().hex}"
+    hyperopt_results = hyperopt(config, dataset=rel_path, output_directory=tmpdir, experiment_name=experiment_name)
     if search_space == "random":
         assert hyperopt_results.experiment_analysis.results_df.shape[0] == RANDOM_SEARCH_SIZE
     else:
@@ -391,10 +379,21 @@ def test_hyperopt_run_hyperopt(csv_filename, search_space, tmpdir, ray_cluster):
     assert isinstance(hyperopt_results, HyperoptResults)
 
     # check for existence of the hyperopt statistics file
-    assert os.path.isfile(os.path.join(tmpdir, "test_hyperopt", HYPEROPT_STATISTICS_FILE_NAME))
+    assert fs_utils.path_exists(os.path.join(tmpdir, experiment_name, HYPEROPT_STATISTICS_FILE_NAME))
+
+
+@pytest.mark.parametrize("fs_protocol,bucket", [private_param(("s3", "ludwig-tests"))], ids=["s3"])
+def test_hyperopt_sync_remote(fs_protocol, bucket, csv_filename, ray_cluster):
+    with remote_tmpdir(fs_protocol, bucket) as tmpdir:
+        with pytest.raises(ValueError) if not _ray200 else contextlib.nullcontext():
+            test_hyperopt_run_hyperopt(
+                csv_filename,
+                "random",
+                tmpdir,
+                ray_cluster,
+            )
 
 
-@pytest.mark.distributed
 def test_hyperopt_with_feature_specific_parameters(csv_filename, tmpdir, ray_cluster):
     input_features = [
         text_feature(name="utterance", reduce_output="sum"),
@@ -446,7 +445,6 @@ def test_hyperopt_with_feature_specific_parameters(csv_filename, tmpdir, ray_clu
             assert input_feature["encoder"]["embedding_size"] in embedding_size_search_space
 
 
-@pytest.mark.distributed
 def test_hyperopt_old_config(csv_filename, tmpdir, ray_cluster):
     old_config = {
         "ludwig_version": "0.4",
@@ -500,7 +498,6 @@ def test_hyperopt_old_config(csv_filename, tmpdir, ray_cluster):
     hyperopt(old_config, dataset=rel_path, output_directory=tmpdir, experiment_name="test_hyperopt")
 
 
-@pytest.mark.distributed
 def test_hyperopt_nested_parameters(csv_filename, tmpdir, ray_cluster):
     config = {
         INPUT_FEATURES: [
@@ -591,7 +588,6 @@ def test_hyperopt_nested_parameters(csv_filename, tmpdir, ray_cluster):
         assert trial_config[TRAINER]["learning_rate"] in {0.7, 0.42}
 
 
-@pytest.mark.distributed
 def test_hyperopt_grid_search_more_than_one_sample(csv_filename, tmpdir, ray_cluster):
     input_features = [
         text_feature(name="utterance", encoder={"reduce_output": "sum"}),

tests/integration_tests/test_remote.py

@@ -1,7 +1,4 @@
-import contextlib
 import os
-import tempfile
-import uuid
 
 import pytest
 import yaml
@@ -11,22 +8,13 @@
 from ludwig.constants import TRAINER
 from ludwig.globals import DESCRIPTION_FILE_NAME
 from ludwig.utils import fs_utils
-from tests.integration_tests.utils import category_feature, generate_data, private_param, sequence_feature
-
-
-@contextlib.contextmanager
-def remote_tmpdir(fs_protocol, bucket):
-    if bucket is None:
-        with tempfile.TemporaryDirectory() as tmpdir:
-            yield f"{fs_protocol}://{tmpdir}"
-        return
-
-    prefix = f"tmp_{uuid.uuid4().hex}"
-    tmpdir = f"{fs_protocol}://{bucket}/{prefix}"
-    try:
-        yield tmpdir
-    finally:
-        fs_utils.delete(tmpdir, recursive=True)
+from tests.integration_tests.utils import (
+    category_feature,
+    generate_data,
+    private_param,
+    remote_tmpdir,
+    sequence_feature,
+)
 
 
 @pytest.mark.parametrize(

tests/integration_tests/utils.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 # ==============================================================================
 
+import contextlib
 import logging
 import multiprocessing
 import os
@@ -40,6 +41,7 @@
 from ludwig.experiment import experiment_cli
 from ludwig.features.feature_utils import compute_feature_hash
 from ludwig.trainers.trainer import Trainer
+from ludwig.utils import fs_utils
 from ludwig.utils.data_utils import read_csv, replace_file_extension
 
 logger = logging.getLogger(__name__)
@@ -872,3 +874,22 @@ def filter(stats):
     finally:
         # Remove results/intermediate data saved to disk
         shutil.rmtree(output_dir, ignore_errors=True)
+
+
+@contextlib.contextmanager
+def remote_tmpdir(fs_protocol, bucket):
+    if bucket is None:
+        with tempfile.TemporaryDirectory() as tmpdir:
+            yield f"{fs_protocol}://{tmpdir}"
+        return
+
+    prefix = f"tmp_{uuid.uuid4().hex}"
+    tmpdir = f"{fs_protocol}://{bucket}/{prefix}"
+    try:
+        yield tmpdir
+    finally:
+        try:
+            fs_utils.delete(tmpdir, recursive=True)
+        except FileNotFoundError as e:
+            logging.info(f"failed to delete remote tempdir, does not exist: {str(e)}")
+            pass