Parallelize data downloading

wikipedia.py

@@ -1039,7 +1039,10 @@ CAT_ALIASES = {
     "zh-yue": ["分类", "分類", "类", "類"],
 }
 
-_BASE_URL_TMPL = "https://dumps.wikimedia.org/{lang}wiki/{date}/"
+# Note that Wikimedia servers have rate limited downloaders and they are capping the number of per-ip connections to 2
+# Their mirror sites do not have this cap: https://dumps.wikimedia.org/mirrors.html
+_HOST = "https://dumps.wikimedia.org"
+_URL_PATH_SEGMENT = "/{lang}wiki/{date}/"
 _INFO_FILE = "dumpstatus.json"
 
 
@@ -1049,14 +1052,19 @@ _VERSION = datasets.Version("4.0.0", "")
 class WikipediaConfig(datasets.BuilderConfig):
     """BuilderConfig for Wikipedia."""
 
-    def __init__(self, language=None, date=None, version=_VERSION, **kwargs):
+    def __init__(self, language=None, date=None, host=_HOST, version=_VERSION, **kwargs):
         """BuilderConfig for Wikipedia.
 
         Args:
-          language: string, the language code for the Wikipedia dump to use.
-          date: string, date of the Wikipedia dump in YYYYMMDD format. A list of
-            available dates can be found at https://dumps.wikimedia.org/enwiki/.
-          **kwargs: keyword arguments forwarded to super.
+            language (str): Language code for the Wikipedia dump to use.
+            date (str): Date of the Wikipedia dump in YYYYMMDD format. A list of
+                available dates can be found at https://dumps.wikimedia.org/enwiki/.
+            host (str, defaults to 'https://dumps.wikimedia.org'): URL of the server that hosts the Wikipedia dump.
+                It defaults to the official Wikimedia host, but this has rate limited downloaders and is capping the
+                number of per-IP connections to 2.
+                To parallelize data download, use a mirror: https://dumps.wikimedia.org/mirrors.html
+
+            **kwargs: Keyword arguments forwarded to super.
         """
         super().__init__(
             name=f"{date}.{language}",
@@ -1066,6 +1074,7 @@ class WikipediaConfig(datasets.BuilderConfig):
         )
         self.date = date
         self.language = language
+        self.host = host.rstrip("/")
 
 
 class Wikipedia(datasets.GeneratorBasedBuilder):
@@ -1090,13 +1099,15 @@ class Wikipedia(datasets.GeneratorBasedBuilder):
         )
 
     def _get_base_url(self):
-        return _BASE_URL_TMPL.format(lang=self.config.language.replace("-", "_"), date=self.config.date)
+        return self.config.host + _URL_PATH_SEGMENT.format(
+            lang=self.config.language.replace("-", "_"), date=self.config.date
+        )
 
     def _split_generators(self, dl_manager):
         # Download dump status info file
         info_url = self._get_base_url() + _INFO_FILE
-        downloaded_files = dl_manager.download_and_extract({"info": info_url})
-        with open(downloaded_files["info"], encoding="utf-8") as f:
+        info_path = dl_manager.download_and_extract(info_url)
+        with open(info_path, encoding="utf-8") as f:
             dump_info = json.load(f)
         multistream_dump_info = dump_info["jobs"]["articlesmultistreamdump"]
         if multistream_dump_info["status"] != "done":
@@ -1115,11 +1126,12 @@ class Wikipedia(datasets.GeneratorBasedBuilder):
             elif ".xml" in fname:
                 total_bytes += info["size"]
                 xml_urls.append(self._get_base_url() + fname)
-        downloaded_paths = dl_manager.download({"index": sorted(index_urls), "xml": sorted(xml_urls)})
+        data_urls = [*zip(sorted(index_urls), sorted(xml_urls))]  # Parallelize data downloading
+        data_paths = dl_manager.download(data_urls)
 
         # Parallelize over concatenated multiple compressed streams (with 100 pages each)
         filepaths, starts, ends = [], [], []
-        for index_path, xml_path in zip(downloaded_paths["index"], downloaded_paths["xml"]):
+        for index_path, xml_path in data_paths:
             index = _extract_index(index_path)
             for start, end in _pairwise(index):
                 filepaths.append(xml_path)