Adds parquet writer (huggingface#103)

* added parquet writer

* nit

* Update src/datatrove/pipeline/writers/parquet.py

Co-authored-by: Mario Šaško <mariosasko777@gmail.com>

* updated test

* nit

---------

Co-authored-by: Mario Šaško <mariosasko777@gmail.com>

Author: guipenedo

src/datatrove/io.py

@@ -23,7 +23,7 @@ def __init__(self, fs, mode: str = "wt", compression: str | None = "infer"):
 
     def get_file(self, filename):
         """
-            Opens file `filename` if it hasn't been opened yet. Otherwise just returns it from the file cache
+            Opens file `filename` if it hasn't been opened yet. Otherwise, just returns it from the file cache
         Args:
           filename: name of the file to open/get if previously opened
 

src/datatrove/pipeline/writers/__init__.py

@@ -1 +1,2 @@
 from .jsonl import JsonlWriter
+from .parquet import ParquetWriter

src/datatrove/pipeline/writers/disk_base.py

@@ -1,7 +1,7 @@
 import dataclasses
 from abc import ABC, abstractmethod
 from string import Template
-from typing import Callable
+from typing import IO, Callable
 
 from datatrove.data import Document, DocumentsPipeline
 from datatrove.io import DataFolderLike, get_datafolder
@@ -31,6 +31,7 @@ def __init__(
         output_filename: str = None,
         compression: str | None = "infer",
         adapter: Callable = None,
+        mode: str = "wt",
     ):
         """
             Base writer block to save data to disk.
@@ -47,7 +48,7 @@ def __init__(
         if self.compression == "gzip" and not output_filename.endswith(".gz"):
             output_filename += ".gz"
         self.output_filename = Template(output_filename)
-        self.output_mg = self.output_folder.get_output_file_manager(mode="wt", compression=compression)
+        self.output_mg = self.output_folder.get_output_file_manager(mode=mode, compression=compression)
         self.adapter = adapter if adapter else _default_adapter
 
     def __enter__(self):
@@ -81,13 +82,13 @@ def _get_output_filename(self, document: Document, rank: int | str = 0, **kwargs
         )
 
     @abstractmethod
-    def _write(self, document: dict, file_handler):
+    def _write(self, document: dict, file_handler: IO, filename: str):
         """
         Main method that subclasses should implement. Receives an adapted (after applying self.adapter) dictionary with data to save to `file_handler`
         Args:
             document: dictionary with the data to save
             file_handler: file_handler where it should be saved
-
+            filename: to use as a key for writer helpers and other data
         Returns:
 
         """
@@ -105,7 +106,7 @@ def write(self, document: Document, rank: int = 0, **kwargs):
 
         """
         output_filename = self._get_output_filename(document, rank, **kwargs)
-        self._write(self.adapter(document), self.output_mg.get_file(output_filename))
+        self._write(self.adapter(document), self.output_mg.get_file(output_filename), output_filename)
         self.stat_update(self._get_output_filename(document, "XXXXX", **kwargs))
         self.stat_update(StatHints.total)
         self.update_doc_stats(document)

src/datatrove/pipeline/writers/jsonl.py

@@ -18,5 +18,5 @@ def __init__(
     ):
         super().__init__(output_folder, output_filename=output_filename, compression=compression, adapter=adapter)
 
-    def _write(self, document: dict, file: IO):
-        file.write(json.dumps(document, ensure_ascii=False) + "\n")
+    def _write(self, document: dict, file_handler: IO, _filename: str):
+        file_handler.write(json.dumps(document, ensure_ascii=False) + "\n")

src/datatrove/pipeline/writers/parquet.py

@@ -0,0 +1,55 @@
+from collections import defaultdict
+from typing import IO, Callable
+
+from datatrove.io import DataFolderLike
+from datatrove.pipeline.writers.disk_base import DiskWriter
+
+
+class ParquetWriter(DiskWriter):
+    default_output_filename: str = "${rank}.parquet"
+    name = "📒 Parquet"
+    _requires_dependencies = ["pyarrow"]
+
+    def __init__(
+        self,
+        output_folder: DataFolderLike,
+        output_filename: str = None,
+        compression: str | None = None,
+        adapter: Callable = None,
+        batch_size: int = 1000,
+    ):
+        super().__init__(output_folder, output_filename, compression, adapter, mode="wb")
+        self._writers = {}
+        self._batches = defaultdict(list)
+        self.batch_size = batch_size
+
+    def _write_batch(self, filename):
+        if not self._batches[filename]:
+            return
+        import pyarrow as pa
+
+        # prepare batch
+        batch = pa.RecordBatch.from_pylist(self._batches.pop(filename))
+        # write batch
+        self._writers[filename].write_batch(batch)
+
+    def _write(self, document: dict, file_handler: IO, filename: str):
+        import pyarrow as pa
+        import pyarrow.parquet as pq
+
+        if filename not in self._writers:
+            self._writers[filename] = pq.ParquetWriter(
+                file_handler, schema=pa.RecordBatch.from_pylist([document]).schema
+            )
+        self._batches[filename].append(document)
+        if len(self._batches[filename]) == self.batch_size:
+            self._write_batch(filename)
+
+    def close(self):
+        for filename in list(self._batches.keys()):
+            self._write_batch(filename)
+        for writer in self._writers.values():
+            writer.close()
+        self._batches.clear()
+        self._writers.clear()
+        super().close()

tests/pipeline/test_parquet_writer.py

@@ -0,0 +1,33 @@
+import shutil
+import tempfile
+import unittest
+
+from datatrove.data import Document
+from datatrove.pipeline.readers.parquet import ParquetReader
+from datatrove.pipeline.writers.parquet import ParquetWriter
+
+from ..utils import require_pyarrow
+
+
+@require_pyarrow
+class TestParquetWriter(unittest.TestCase):
+    def setUp(self):
+        # Create a temporary directory
+        self.tmp_dir = tempfile.mkdtemp()
+        self.addCleanup(shutil.rmtree, self.tmp_dir)
+
+    def test_write(self):
+        data = [
+            Document(text=text, id=str(i), metadata={"somedata": 2 * i, "somefloat": i * 0.4, "somestring": "hello"})
+            for i, text in enumerate(["hello", "text2", "more text"])
+        ]
+        with ParquetWriter(output_folder=self.tmp_dir, batch_size=2) as w:
+            for doc in data:
+                w.write(doc)
+        reader = ParquetReader(self.tmp_dir)
+        c = 0
+        for read_doc, original in zip(reader(), data):
+            read_doc.metadata.pop("file_path", None)
+            assert read_doc == original
+            c += 1
+        assert c == len(data)