changes for feat:add-parquet-support PR review (#460)

README.md

@@ -643,14 +643,18 @@ The `overwrite` mode will delete the existing data and start from fresh.
 </details>
 
 <details>
-  <summary> ✅ Index parquet datasets</summary>
+  <summary> ✅ Stream parquet datasets</summary>
 &nbsp;
 
-If your dataset is already in Parquet format, you can index it directly and use it with StreamingDataset & DataLoader.
+You can stream Parquet datasets directly without the need to convert them into the LitData optimized binary format.
+
+If your dataset is already in Parquet format, you can index and use it with StreamingDataset and DataLoader for efficient streaming.
 
 Assumption:
 Your dataset directory contains one or more Parquet files.
 
+- **Index Parquet dataset**:
+
 ```python
 import litdata as ld
 
@@ -659,6 +663,8 @@ pq_data_uri = "gs://deep-litdata-parquet/my-parquet-data"
 ld.index_parquet_dataset(pq_data_uri)
 ```
 
+- **Stream the dataset with `StreamingDataset` and `ParquetLoader`**
+
 When using a Streaming Dataset, ensure you use `ParquetLoader`:
 
 ```python

src/litdata/constants.py

@@ -36,7 +36,7 @@
 _AZURE_STORAGE_AVAILABLE = RequirementCache("azure.storage.blob")
 _TQDM_AVAILABLE = RequirementCache("tqdm")
 _LIGHTNING_SDK_AVAILABLE = RequirementCache("lightning_sdk")
-_POLARS_AVAILABLE = RequirementCache("polars")
+_POLARS_AVAILABLE = RequirementCache("polars>1.0.0")
 _DEBUG = bool(int(os.getenv("DEBUG", "1")))
 
 _MAX_WAIT_TIME = int(os.getenv("MAX_WAIT_TIME", "120"))

src/litdata/streaming/item_loader.py

@@ -423,7 +423,10 @@ def encode_data(cls, data: List[bytes], _: List[int], flattened: List[Any]) -> T
 class ParquetLoader(BaseItemLoader):
     def __init__(self) -> None:
         if not _POLARS_AVAILABLE:
-            raise ModuleNotFoundError("Please, run: `pip install polars`")
+            raise ModuleNotFoundError(
+                "You are using the Parquet item loader, which depends on `Polars > 1.0.0`.",
+                "Please, run: `pip install polars>1.0.0`",
+            )
         self._chunk_filepaths: Dict[str, bool] = {}
 
     def setup(
@@ -458,7 +461,10 @@ def generate_intervals(self) -> List[Interval]:
 
     def pre_load_chunk(self, chunk_index: int, chunk_filepath: str) -> None:
         """Logic to load the chunk in background to gain some time."""
-        pass
+        import polars as pl
+
+        if chunk_filepath not in self._df:
+            self._df[chunk_filepath] = pl.scan_parquet(chunk_filepath).collect()
 
     def load_item_from_chunk(
         self,

tests/streaming/test_writer.py

@@ -15,6 +15,7 @@
 import os
 import random
 import sys
+from collections import OrderedDict
 
 import numpy as np
 import pytest
@@ -278,14 +279,16 @@ def test_parquet_index_write(tmpdir):
 
     os.mkdir(os.path.join(tmpdir, "data"))
 
+    pq_data = OrderedDict(
+        {
+            "name": ["Tom", "Jerry", "Micky", "Oggy", "Doraemon"],
+            "weight": [57.9, 72.5, 53.6, 83.1, 69.4],  # (kg)
+            "height": [1.56, 1.77, 1.65, 1.75, 1.63],  # (m)
+        }
+    )
+
     for i in range(5):
-        df = pl.DataFrame(
-            {
-                "name": ["Tom", "Jerry", "Micky", "Oggy", "Doraemon"],
-                "weight": [57.9, 72.5, 53.6, 83.1, 69.4],  # (kg)
-                "height": [1.56, 1.77, 1.65, 1.75, 1.63],  # (m)
-            }
-        )
+        df = pl.DataFrame(pq_data)
         file_path = os.path.join(tmpdir, "data", f"tmp-{i}.parquet")
         df.write_parquet(file_path)
 
@@ -307,3 +310,10 @@ def test_parquet_index_write(tmpdir):
     ds = StreamingDataset(os.path.join(tmpdir, "data"), item_loader=ParquetLoader())
 
     assert len(ds) == 25  # 5 datasets for 5 loops
+
+    for i, _ds in enumerate(ds):
+        idx = i % 5
+        assert len(_ds) == 3
+        assert _ds[0] == pq_data["name"][idx]
+        assert _ds[1] == pq_data["weight"][idx]
+        assert _ds[2] == pq_data["height"][idx]