Add support for Dataflux Iterable Dataset (#17)

* add Dataflux Iterable Dataset

* rename to DataFluxIterableDataset

* add test case for multi-worker setup

* update license header year

* update license header year #2

* Address comments

Author: bernardhan33

dataflux_client_python

@@ -0,0 +1,171 @@
+"""
+ Copyright 2024 Google LLC
+
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+
+      https://www.apache.org/licenses/LICENSE-2.0
+
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License.
+ """
+
+import os
+import math
+import logging
+
+from torch.utils import data
+from google.cloud import storage
+from google.api_core.client_info import ClientInfo
+
+import dataflux_core
+
+
+class Config:
+    """Customizable configuration to the DataFluxIterableDataset.
+
+    Attributes:
+        sort_listing_results: A boolean flag indicating if data listing results
+            will be alphabetically sorted. Default to False.
+
+        max_composite_object_size: An integer indicating a cap for the maximum
+            size of the composite object in bytes. Default to 100000000 = 100 MiB.
+
+        num_processes: The number of processes to be used in the Dataflux algorithms.
+            Default to the number of CPUs from the running environment.
+
+        prefix: The prefix that is used to list the objects in the bucket with.
+            The default is None which means it will list all the objects in the bucket.
+
+        max_listing_retries: An integer indicating the maximum number of retries
+        to attempt in case of any Python multiprocessing errors during
+        GCS objects listing. Default to 3.
+    """
+
+    def __init__(
+        self,
+        sort_listing_results: bool = False,
+        max_composite_object_size: int = 100000000,
+        num_processes: int = os.cpu_count(),
+        prefix: str = None,
+        max_listing_retries: int = 3,
+    ):
+        self.sort_listing_results = sort_listing_results
+        self.max_composite_object_size = max_composite_object_size
+        self.num_processes = num_processes
+        self.prefix = prefix
+        self.max_listing_retries = max_listing_retries
+
+
+class DataFluxIterableDataset(data.IterableDataset):
+    def __init__(
+        self,
+        project_name,
+        bucket_name,
+        config=Config(),
+        data_format_fn=lambda data: data,
+        storage_client=None,
+    ):
+        """Initializes the DataFluxIterableDataset.
+
+        The initialization sets up the needed configuration and runs data
+        listing using the Dataflux algorithm.
+
+        Args:
+            project_name: The name of the GCP project.
+            bucket_name: The name of the GCS bucket that holds the objects to compose.
+                The Dataflux download algorithm uploads the the composed object to this bucket too.
+            destination_blob_name: The name of the composite object to be created.
+            config: A dataflux_iterable_dataset.Config object that includes configuration
+                customizations. If not specified, a default config with default parameters is created.
+            data_format_fn: A function that formats the downloaded bytes to the desired format.
+                If not specified, the default formatting function leaves the data as-is.
+            storage_client: The google.cloud.storage.Client object initiated with sufficient permission
+                to access the project and the bucket. If not specified, it will be created
+                during initialization.
+        """
+        super().__init__()
+        self.storage_client = storage_client
+        if not storage_client:
+            self.storage_client = storage.Client(
+                project=project_name,
+                client_info=ClientInfo(user_agent="dataflux/0.0"),
+            )
+        self.project_name = project_name
+        self.bucket_name = bucket_name
+        self.data_format_fn = data_format_fn
+        self.config = config
+        self.dataflux_download_optimization_params = (
+            dataflux_core.download.DataFluxDownloadOptimizationParams(
+                max_composite_object_size=self.config.max_composite_object_size
+            )
+        )
+
+        self.objects = self._list_GCS_blobs_with_retry()
+
+    def __iter__(self):
+        worker_info = data.get_worker_info()
+        if worker_info is None:
+            # Single-process data loading.
+            yield from [
+                self.data_format_fn(bytes_content)
+                for bytes_content in dataflux_core.download.dataflux_download_lazy(
+                    project_name=self.project_name,
+                    bucket_name=self.bucket_name,
+                    objects=self.objects,
+                    storage_client=self.storage_client,
+                    dataflux_download_optimization_params=self.dataflux_download_optimization_params,
+                )
+            ]
+        else:
+            # Multi-process data loading. Split the workload among workers.
+            # Ref: https://pytorch.org/docs/stable/data.html#torch.utils.data.IterableDataset.
+            per_worker = int(
+                math.ceil(len(self.objects) / float(worker_info.num_workers))
+            )
+            worker_id = worker_info.id
+            start = worker_id * per_worker
+            end = min(start + per_worker, len(self.objects))
+            yield from [
+                self.data_format_fn(bytes_content)
+                for bytes_content in dataflux_core.download.dataflux_download_lazy(
+                    project_name=self.project_name,
+                    bucket_name=self.bucket_name,
+                    objects=self.objects[start:end],
+                    storage_client=self.storage_client,
+                    dataflux_download_optimization_params=self.dataflux_download_optimization_params,
+                )
+            ]
+
+    def _list_GCS_blobs_with_retry(self):
+        """Retries Dataflux Listing upon exceptions, up to the retries defined in self.config."""
+        error = None
+        listed_objects = []
+        for _ in range(self.config.max_listing_retries):
+            try:
+                listed_objects = dataflux_core.fast_list.ListingController(
+                    max_parallelism=self.config.num_processes,
+                    project=self.project_name,
+                    bucket=self.bucket_name,
+                    sort_results=self.config.sort_listing_results,
+                    prefix=self.config.prefix,
+                ).run()
+            except Exception as e:
+                logging.error(
+                    f"exception {str(e)} caught running Dataflux fast listing."
+                )
+                error = e
+                continue
+
+            # No exception -- we can immediately return the listed objects.
+            else:
+                return listed_objects
+
+        # Did not break the for loop, therefore all attempts
+        # raised an exception.
+        else:
+            raise error

dataflux_pytorch/dataflux_iterable_dataset.py

@@ -77,7 +77,7 @@ def __init__(
         Args:
             project_name: The name of the GCP project.
             bucket_name: The name of the GCS bucket that holds the objects to compose.
-                The function uploads the the composed object to this bucket too.
+                The Dataflux download algorithm uploads the the composed object to this bucket too.
             destination_blob_name: The name of the composite object to be created.
             config: A dataflux_mapstyle_dataset.Config object that includes configuration
                 customizations. If not specified, a default config with default parameters is created.