[MLFlowObjectStore] [1/2] Base implementation for MLFlowObjectStore

composer/utils/__init__.py

@@ -21,8 +21,9 @@
 from composer.utils.iter_helpers import IteratorFileStream, ensure_tuple, map_collection
 from composer.utils.misc import (create_interval_scheduler, get_free_tcp_port, is_model_deepspeed, is_model_fsdp,
                                  is_notebook, model_eval_mode, using_torch_2)
-from composer.utils.object_store import (GCSObjectStore, LibcloudObjectStore, ObjectStore, ObjectStoreTransientError,
-                                         OCIObjectStore, S3ObjectStore, SFTPObjectStore, UCObjectStore)
+from composer.utils.object_store import (GCSObjectStore, LibcloudObjectStore, MLFlowObjectStore, ObjectStore,
+                                         ObjectStoreTransientError, OCIObjectStore, S3ObjectStore, SFTPObjectStore,
+                                         UCObjectStore)
 from composer.utils.retrying import retry
 from composer.utils.string_enum import StringEnum
 
@@ -44,6 +45,7 @@
     'OCIObjectStore',
     'GCSObjectStore',
     'UCObjectStore',
+    'MLFlowObjectStore',
     'MissingConditionalImportError',
     'import_object',
     'is_model_deepspeed',

composer/utils/object_store/__init__.py

@@ -5,13 +5,14 @@
 
 from composer.utils.object_store.gcs_object_store import GCSObjectStore
 from composer.utils.object_store.libcloud_object_store import LibcloudObjectStore
+from composer.utils.object_store.mlflow_object_store import MLFlowObjectStore
 from composer.utils.object_store.object_store import ObjectStore, ObjectStoreTransientError
 from composer.utils.object_store.oci_object_store import OCIObjectStore
 from composer.utils.object_store.s3_object_store import S3ObjectStore
 from composer.utils.object_store.sftp_object_store import SFTPObjectStore
 from composer.utils.object_store.uc_object_store import UCObjectStore
 
 __all__ = [
-    'ObjectStore', 'ObjectStoreTransientError', 'LibcloudObjectStore', 'S3ObjectStore', 'SFTPObjectStore',
-    'OCIObjectStore', 'GCSObjectStore', 'UCObjectStore'
+    'ObjectStore', 'ObjectStoreTransientError', 'LibcloudObjectStore', 'MLFlowObjectStore', 'S3ObjectStore',
+    'SFTPObjectStore', 'OCIObjectStore', 'GCSObjectStore', 'UCObjectStore'
 ]

composer/utils/object_store/object_store.py

@@ -87,7 +87,7 @@ def upload_object(self,
 
         Args:
             object_name (str): Object name (where object will be stored in the container)
-            filename (str | pathlib.Path): Path the the object on disk
+            filename (str | pathlib.Path): Path to the object on disk
             callback ((int, int) -> None, optional): If specified, the callback is periodically called with the number of bytes
                 uploaded and the total size of the object being uploaded.
             **kwargs: other arguments to the upload object function are supported
@@ -133,6 +133,7 @@ def download_object(
                 downloaded and the total size of the object.
 
         Raises:
+            FileExistsError: If ``filename`` already exists and ``overwrite`` is ``False``.
             FileNotFoundError: If the file was not found in the object store.
             ObjectStoreTransientError: If there was a transient connection issue with downloading the object.
         """

composer/utils/object_store/uc_object_store.py

@@ -53,7 +53,7 @@ def __init__(self, path: str) -> None:
         try:
             from databricks.sdk import WorkspaceClient
         except ImportError as e:
-            raise MissingConditionalImportError('databricks', conda_package='databricks-sdk>=0.8.0,<1.0') from e
+            raise MissingConditionalImportError('databricks', conda_package='databricks-sdk>=0.15.0,<1.0') from e
 
         try:
             self.client = WorkspaceClient()
@@ -167,7 +167,7 @@ def download_object(self,
         try:
             from databricks.sdk.core import DatabricksError
             try:
-                with self.client.files.download(self._get_object_path(object_name)).contents as resp:
+                with self.client.files.download(self._get_object_path(object_name)).contents as resp:  # pyright: ignore
                     with open(tmp_path, 'wb') as f:
                         # Chunk the data into multiple blocks of 64MB to avoid
                         # OOMs when downloading really large files
@@ -199,11 +199,15 @@ def get_object_size(self, object_name: str) -> int:
 
         Raises:
             FileNotFoundError: If the file was not found in the object store.
+            IsADirectoryError: If the object is a directory, not a file.
         """
         from databricks.sdk.core import DatabricksError
         try:
             file_info = self.client.files.get_status(self._get_object_path(object_name))
-            return file_info.file_size
+            if file_info.is_dir:
+                raise IsADirectoryError(f'{object_name} is a UC directory, not a file.')
+
+            return file_info.file_size  # pyright: ignore
         except DatabricksError as e:
             _wrap_errors(self.get_uri(object_name), e)
 
@@ -231,6 +235,6 @@ def list_objects(self, prefix: Optional[str]) -> List[str]:
                                              path=self._UC_VOLUME_LIST_API_ENDPOINT,
                                              data=data,
                                              headers={'Source': 'mosaicml/composer'})
-            return [f['path'] for f in resp.get('files', []) if not f['is_dir']]
+            return [f['path'] for f in resp.get('files', []) if not f['is_dir']]  # pyright: ignore
         except DatabricksError as e:
             _wrap_errors(self.get_uri(prefix), e)

setup.py

@@ -70,7 +70,8 @@ def package_files(prefix: str, directory: str, extension: str):
         break
     else:
         assert end != -1, 'there should be a balanced number of start and ends'
-        long_description = long_description[:start] + long_description[end + len(end_tag):]
+        long_description = long_description[:start] + \
+            long_description[end + len(end_tag):]
 
 install_requires = [
     'pyyaml>=6.0,<7',
@@ -223,7 +224,7 @@ def package_files(prefix: str, directory: str, extension: str):
 ]
 
 extra_deps['mlflow'] = [
-    'mlflow>=2.9.0,<3.0',
+    'mlflow>=2.9.2,<3.0',
 ]
 
 extra_deps['pandas'] = ['pandas>=2.0.0,<3.0']

tests/utils/object_store/object_store_settings.py

@@ -14,8 +14,8 @@
 
 import composer.utils.object_store
 import composer.utils.object_store.sftp_object_store
-from composer.utils.object_store import (GCSObjectStore, LibcloudObjectStore, ObjectStore, OCIObjectStore,
-                                         S3ObjectStore, SFTPObjectStore, UCObjectStore)
+from composer.utils.object_store import (GCSObjectStore, LibcloudObjectStore, MLFlowObjectStore, ObjectStore,
+                                         OCIObjectStore, S3ObjectStore, SFTPObjectStore, UCObjectStore)
 from composer.utils.object_store.sftp_object_store import SFTPObjectStore
 from tests.common import get_module_subclasses
 
@@ -56,8 +56,9 @@
 object_stores = [
     pytest.param(x, marks=_object_store_marks[x], id=x.__name__)
     for x in get_module_subclasses(composer.utils.object_store, ObjectStore)
-    # Note: OCI, GCS and UC have their own test suite, so they are exempt from being included in this one.``
-    if not issubclass(x, OCIObjectStore) and not issubclass(x, GCSObjectStore) and not issubclass(x, UCObjectStore)
+    # Note: OCI, GCS, UC, and MLFlow have their own test suite, so they are exempt from being included in this one.``
+    if not issubclass(x, OCIObjectStore) and not issubclass(x, GCSObjectStore) and not issubclass(x, UCObjectStore) and
+    not issubclass(x, MLFlowObjectStore)
 ]