Add support for parallelizing processing parquet files across workers…

… and nodes. (#19400)

requirements/data/test.txt

@@ -5,3 +5,5 @@ pytest-timeout ==2.1.0
 pytest-rerunfailures ==12.0
 pytest-random-order ==1.1.0
 viztracer
+pyarrow
+polars

src/lightning/data/processing/image.py

@@ -0,0 +1,47 @@
+import io
+from typing import Optional, Tuple
+
+from lightning_utilities.core.imports import RequirementCache
+
+_HTTPX_AVAILABLE = RequirementCache("httpx")
+
+# Credit to the https://github.com/rom1504/img2dataset Github repo
+# The code was taken from there. It has a MIT License.
+
+def _download_image(
+    url: str,
+    timeout: int = 10,
+    user_agent_token: str = "pytorch-lightning",
+) -> Tuple[Optional[io.BytesIO], Optional[Exception]]:
+    """Download an image with urllib."""
+    url
+    img_stream = None
+    user_agent_string = "Mozilla/5.0 (X11; Ubuntu; Linux x86_64; rv:72.0) Gecko/20100101 Firefox/72.0"
+    if user_agent_token:
+        user_agent_string += f" (compatible; {user_agent_token}; +https://github.com/Lightning-AI/pytorch-lightning)"
+    import httpx
+
+    try:
+        with httpx.Client(http2=True) as client:
+            r = client.get(url, headers={"User-Agent": user_agent_string}, timeout=timeout)
+            img_stream = io.BytesIO(r.read())
+        return img_stream, None
+    except Exception as err:  # pylint: disable=broad-except
+        if img_stream is not None:
+            img_stream.close()
+        return None, err
+
+
+def download_image(
+    url: str,
+    retries: int = 0,
+    timeout: int = 10,
+    user_agent_token: str = "pytorch-lightning",
+) -> Tuple[Optional[io.BytesIO], Optional[Exception]]:
+    if not _HTTPX_AVAILABLE:
+        raise ModuleNotFoundError("Please, run: `pip install httpx`.")
+    for _ in range(retries + 1):
+        img_stream, err = _download_image(url, timeout, user_agent_token)
+        if img_stream is not None:
+            return img_stream, err
+    return None, err

src/lightning/data/processing/readers.py

@@ -0,0 +1,131 @@
+import os
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import Any, List, Optional
+
+from lightning_utilities.core.imports import RequirementCache
+
+from lightning.data.streaming.shuffle import _associate_chunks_and_internals_to_ranks
+from lightning.data.utilities.env import _DistributedEnv
+
+_POLARS_AVAILABLE = RequirementCache("polars")
+_PYARROW_AVAILABLE = RequirementCache("pyarrow")
+
+
+class BaseReader(ABC):
+
+    def get_num_nodes(self) -> int:
+        return int(os.getenv("DATA_OPTIMIZER_NUM_NODES", 1))
+
+    def get_node_rank(self) -> int:
+        return int(os.getenv("DATA_OPTIMIZER_NODE_RANK", 0))
+
+    @abstractmethod
+    def items_to_workers(self, items: List[Any], num_workers: int) -> List[List[Any]]:
+        """This method is meant to convert the items provided by the users into items to be processed by the
+        workers."""
+        pass
+
+    @abstractmethod
+    def read(self, item: Any) -> Any:
+        """Read the data associated to an item."""
+        pass
+
+
+@dataclass
+class ParquetSlice:
+    """Keep track of a parquet file slice with its filepath, start and end."""
+    filepath: str
+    start: int
+    end: int
+
+
+class ParquetReader(BaseReader):
+
+    def __init__(self, num_rows: Optional[int] = 2048, to_pandas: bool = True) -> None:
+        self.num_rows = num_rows
+        self.to_pandas = to_pandas
+
+        if not _PYARROW_AVAILABLE or not _POLARS_AVAILABLE:
+            raise ModuleNotFoundError("Please, run: `pip install pyarrow polars`")
+
+    def _get_num_rows(self, path: str) -> int:
+        if _PYARROW_AVAILABLE:
+            import pyarrow.dataset as ds
+            df = ds.dataset(path).scanner()
+            return df.count_rows()
+
+        # FIXED: There is a bug in polars. This leads to read_parquet to hang.
+        if _POLARS_AVAILABLE:
+            import polars as pol
+            df = pol.scan_parquet(path)
+            num_rows = df.select(pol.len()).collect().item()
+            return num_rows
+
+        raise RuntimeError("Please, install either pyarrow or polars.")
+
+    def read(self, item: ParquetSlice) -> Any:
+        if _POLARS_AVAILABLE:
+            import polars as pol
+            df = pol.scan_parquet(item.filepath).slice(item.start, item.end).collect()
+
+            if self.to_pandas:
+                df = df.to_pandas()
+
+            return df
+
+        if _PYARROW_AVAILABLE:
+            import pyarrow.dataset as ds
+
+            df = ds.dataset(item.filepath).scanner()
+
+            df = df.take([item.start, item.end])
+
+            if self.to_pandas:
+                df.to_pandas()
+
+            return df
+
+        raise RuntimeError("Please, install either pyarrow or polars.")
+
+
+    def items_to_workers(self, items: Any, num_workers: int) -> List[List[ParquetSlice]]:
+        intervals = [(0, self._get_num_rows(item)) for item in items]
+
+        world_size = self.get_num_nodes() * num_workers
+        node_rank = self.get_node_rank()
+
+        fake_distributed_env = _DistributedEnv(world_size, 0, self.get_num_nodes())
+        parquet_indexes_per_worker, p_slices_per_worker = _associate_chunks_and_internals_to_ranks(
+            fake_distributed_env, list(range(len(items))), intervals, False)
+
+        workers_user_items: List[List[ParquetSlice]] = [[] for _ in range(num_workers)]
+
+        iterator = enumerate(zip(parquet_indexes_per_worker, p_slices_per_worker))
+
+        node_start = node_rank * num_workers
+        node_end = (node_rank + 1) * num_workers
+
+        for worker_idx, (parquet_indexes, p_slices) in iterator:
+            if node_start <= worker_idx < node_end:
+                if self.num_rows:
+                    workers_user_items[worker_idx % num_workers].extend([
+                        ParquetSlice(
+                            items[parquet_index], p_slice_start, p_slice_start + self.num_rows
+                            if p_slice[1] > (p_slice_start + self.num_rows) else
+                            p_slice[1]
+                        )
+                        for parquet_index, p_slice in zip(parquet_indexes, p_slices)
+                        for p_slice_start in range(p_slice[0], p_slice[1] + self.num_rows, self.num_rows)
+                        if p_slice_start < p_slice[1]
+                    ])
+                else:
+                    workers_user_items[worker_idx % num_workers].extend([
+                        ParquetSlice(items[parquet_index], *p_slice)
+                        for parquet_index, p_slice in zip(parquet_indexes, p_slices)
+                    ])
+
+        assert len(workers_user_items) == num_workers
+        assert all(len(w) for w in workers_user_items)
+
+        return workers_user_items

src/lightning/data/streaming/data_processor.py

@@ -20,6 +20,7 @@
 from tqdm.auto import tqdm as _tqdm
 
 from lightning import seed_everything
+from lightning.data.processing.readers import BaseReader
 from lightning.data.streaming import Cache
 from lightning.data.streaming.cache import Dir
 from lightning.data.streaming.client import S3Client
@@ -158,8 +159,9 @@ def _download_data_target(input_dir: Dir, cache_dir: str, queue_in: Queue, queue
                         s3.client.download_fileobj(obj.netloc, obj.path.lstrip("/"), f)
 
                 elif os.path.isfile(path):
-                    os.makedirs(os.path.dirname(local_path), exist_ok=True)
-                    shutil.copyfile(path, local_path)
+                    if not path.startswith("/teamspace/studios/this_studio"):
+                        os.makedirs(os.path.dirname(local_path), exist_ok=True)
+                        shutil.copyfile(path, local_path)
                 else:
                     raise ValueError(f"The provided {input_dir.url} isn't supported.")
 
@@ -340,6 +342,7 @@ def __init__(
         num_downloaders: int,
         num_uploaders: int,
         remove: bool,
+        reader: Optional[BaseReader] = None,
     ) -> None:
         """The BaseWorker is responsible to process the user data."""
         self.worker_index = worker_index
@@ -353,6 +356,7 @@ def __init__(
         self.num_downloaders = num_downloaders
         self.num_uploaders = num_uploaders
         self.remove = remove
+        self.reader = reader
         self.paths: List[List[str]] = []
         self.remover: Optional[Process] = None
         self.downloaders: List[Process] = []
@@ -433,7 +437,7 @@ def _loop(self) -> None:
                 self.progress_queue.put((self.worker_index, self._counter))
                 self._last_time = time()
 
-            if self.remove and self.input_dir.path is not None:
+            if self.remove and self.input_dir.path is not None and self.reader is None:
                 self.remove_queue.put(self.paths[index])
 
             try:
@@ -476,7 +480,7 @@ def _try_upload(self, data: Optional[Union[str, Tuple[str, str]]]) -> None:
         self.to_upload_queues[self._counter % self.num_uploaders].put(data)
 
     def _collect_paths(self) -> None:
-        if self.input_dir.path is None:
+        if self.input_dir.path is None or self.reader is not None:
             for index in range(len(self.items)):
                 self.ready_to_process_queue.put(index)
             for _ in range(self.num_downloaders):
@@ -513,7 +517,7 @@ def is_path(element: Any) -> bool:
             paths = []
             for index, path in indexed_paths.items():
                 paths.append(path)
-                if self.input_dir:
+                if self.input_dir and not self.input_dir.path.startswith("/teamspace/studios/this_studio"):
                     path = path.replace(self.input_dir.path, self.cache_data_dir)
                 flattened_item[index] = path
 
@@ -525,8 +529,9 @@ def is_path(element: Any) -> bool:
         self.items = items
 
     def _start_downloaders(self) -> None:
-        if self.input_dir.path is None:
+        if self.input_dir.path is None or self.reader is not None:
             return
+
         for _ in range(self.num_downloaders):
             to_download_queue: Queue = Queue()
             p = Process(
@@ -583,7 +588,7 @@ def _start_uploaders(self) -> None:
 
     def _handle_data_chunk_recipe(self, index: int) -> None:
         try:
-            self._current_item = self.items[index]
+            self._current_item = self.items[index] if self.reader is None else self.reader.read(self.items[index])
             item_data_or_generator = self.data_recipe.prepare_item(self._current_item)
             if isinstance(item_data_or_generator, types.GeneratorType):
                 for item_data in item_data_or_generator:
@@ -596,7 +601,7 @@ def _handle_data_chunk_recipe(self, index: int) -> None:
                 self._try_upload(chunk_filepath)
                 self._index_counter += 1
         except Exception as e:
-            raise RuntimeError(f"Failed processing {self._current_item}") from e
+            raise RuntimeError(f"Failed processing {self.items[index]}") from e
 
     def _handle_data_chunk_recipe_end(self) -> None:
         chunks_filepaths = self.cache.done()
@@ -609,7 +614,8 @@ def _handle_data_chunk_recipe_end(self) -> None:
     def _handle_data_transform_recipe(self, index: int) -> None:
         # Don't use a context manager to avoid deleting files that are being uploaded.
         output_dir = tempfile.mkdtemp()
-        item_data = self.data_recipe.prepare_item(self.items[index], str(output_dir), len(self.items) - 1 == index)
+        item = self.items[index] if self.reader is None else self.reader.read(self.items[index])
+        item_data = self.data_recipe.prepare_item(item, str(output_dir), len(self.items) - 1 == index)
         if item_data is not None:
             raise ValueError(
                 "When using a `DataTransformRecipe`, the `prepare_item` shouldn't return anything."
@@ -792,6 +798,7 @@ def __init__(
         random_seed: Optional[int] = 42,
         reorder_files: bool = True,
         weights: Optional[List[int]] = None,
+        reader: Optional[BaseReader] = None,
     ):
         """The `DatasetOptimiser` provides an efficient way to process data across multiple machine into chunks to make
         training faster.
@@ -809,6 +816,7 @@ def __init__(
                 Set this to ``False`` if the order in which samples are processed should be preserved.
             weights: Provide a list of weights associated to the inputs.
                 This is used to evenly split the work among the workers.
+            reader: Map the inputs to worker inputs and provides a read method to read a slice of the data.
 
         """
         self.input_dir = _resolve_dir(input_dir)
@@ -825,6 +833,10 @@ def __init__(
         self.stop_queues: List[Queue] = []
         self.reorder_files = reorder_files
         self.weights = weights
+        self.reader = reader
+
+        if self.reader is not None and self.weights is not None:
+            raise ValueError("Either the reader or the weights needs to be defined.")
 
         # Ensure the input dir is the same across all nodes
         self.input_dir = broadcast_object("input_dir", self.input_dir)
@@ -853,7 +865,10 @@ def run(self, data_recipe: DataRecipe) -> None:
         if not isinstance(user_items, list):
             raise ValueError("The `prepare_structure` should return a list of item metadata.")
 
-        if self.weights is not None:
+        if self.reader:
+            workers_user_items = self.reader.items_to_workers(user_items, self.num_workers)
+
+        elif self.weights is not None:
             if len(self.weights) != len(user_items):
                 raise ValueError("The provided weights length should match the inputs' length.")
             workers_user_items = _map_items_to_workers_weighted(
@@ -880,7 +895,7 @@ def run(self, data_recipe: DataRecipe) -> None:
 
         self._cleanup_cache()
 
-        print(f"Starting {self.num_workers} workers")
+        print(f"Starting {self.num_workers} workers with {num_items} items.")
 
         if self.input_dir is None and self.src_resolver is not None and self.input_dir:
             self.input_dir = self.src_resolver(self.input_dir)
@@ -988,6 +1003,7 @@ def _create_process_workers(self, data_recipe: DataRecipe, workers_user_items: L
                 self.num_downloaders,
                 self.num_uploaders,
                 self.delete_cached_files,
+                self.reader,
             )
             worker.start()
             workers.append(worker)

src/lightning/data/streaming/functions.py

@@ -22,6 +22,7 @@
 
 import torch
 
+from lightning.data.processing.readers import BaseReader
 from lightning.data.streaming.constants import _TORCH_GREATER_EQUAL_2_1_0
 from lightning.data.streaming.data_processor import DataChunkRecipe, DataProcessor, DataTransformRecipe
 from lightning.data.streaming.resolver import (
@@ -157,6 +158,7 @@ def map(
     num_downloaders: Optional[int] = None,
     reorder_files: bool = True,
     error_when_not_empty: bool = False,
+    reader: Optional[BaseReader] = None,
 ) -> None:
     """This function map a callbable over a collection of files possibly in a distributed way.
 
@@ -203,6 +205,7 @@ def map(
             num_downloaders=num_downloaders,
             reorder_files=reorder_files,
             weights=weights,
+            reader=reader,
         )
         return data_processor.run(LambdaDataTransformRecipe(fn, inputs))
     return _execute(
@@ -225,6 +228,7 @@ def optimize(
     machine: Optional[str] = None,
     num_downloaders: Optional[int] = None,
     reorder_files: bool = True,
+    reader: Optional[BaseReader] = None,
 ) -> None:
     """This function converts a dataset into chunks possibly in a distributed way.
 
@@ -274,6 +278,7 @@ def optimize(
             fast_dev_run=fast_dev_run,
             num_downloaders=num_downloaders,
             reorder_files=reorder_files,
+            reader=reader,
         )
         return data_processor.run(
             LambdaDataChunkRecipe(