Make Language.pipe workers exit cleanly

Also warn when any worker exited with a non-zero exit code and modify
test to ensure that workers exit cleanly by default.

spacy/errors.py

@@ -220,6 +220,7 @@ class Warnings(metaclass=ErrorsWithCodes):
             "key attribute for vectors, configure it through Vectors(attr=) or "
             "'spacy init vectors --attr'")
     W126 = ("These keys are unsupported: {unsupported}")
+    W127 = ("Not all `Language.pipe` worker processes completed successfully")
 
 
 class Errors(metaclass=ErrorsWithCodes):

spacy/language.py

@@ -1730,6 +1730,9 @@ def prepare_input(
             for proc in procs:
                 proc.join()
 
+            if not all(proc.exitcode == 0 for proc in procs):
+                warnings.warn(Warnings.W127)
+
     def _link_components(self) -> None:
         """Register 'listeners' within pipeline components, to allow them to
         effectively share weights.
@@ -2350,6 +2353,7 @@ def _apply_pipes(
             if isinstance(texts_with_ctx, _WorkDoneSentinel):
                 sender.close()
                 receiver.close()
+                return
 
             docs = (
                 ensure_doc(doc_like, context) for doc_like, context in texts_with_ctx
@@ -2375,6 +2379,7 @@ def _apply_pipes(
             # stop processing.
             sender.close()
             receiver.close()
+            return
 
 
 class _Sender:

spacy/tests/test_language.py

@@ -1,4 +1,5 @@
 import itertools
+import warnings
 import logging
 from unittest import mock
 
@@ -738,9 +739,13 @@ def test_pass_doc_to_pipeline(nlp, n_process):
     assert doc.text == texts[0]
     assert len(doc.cats) > 0
     if isinstance(get_current_ops(), NumpyOps) or n_process < 2:
-        docs = nlp.pipe(docs, n_process=n_process)
-        assert [doc.text for doc in docs] == texts
-        assert all(len(doc.cats) for doc in docs)
+        # Catch warnings to ensure that all worker processes exited
+        # succesfully.
+        with warnings.catch_warnings():
+            warnings.simplefilter("error")
+            docs = nlp.pipe(docs, n_process=n_process)
+            assert [doc.text for doc in docs] == texts
+            assert all(len(doc.cats) for doc in docs)
 
 
 def test_invalid_arg_to_pipeline(nlp):