fix(stream_io): Finalize temporary files before the world ends

Previously, automatically uploaded temporary files for the s3 streams
from `open_stream` were not guaranteed to trigger their upload and
cleanup before the interpreter began shutting down.
This change guarantees that the upload and deletion will happen at the
soonest of either:
1. The stream being closed,
2. The stream being garbage collected, or
3. The program ending, but before the interpreter self-destructs.

Author: Eta0

tensorizer/stream_io.py

@@ -1,11 +1,12 @@
 import functools
+import io
 import logging
 import os
 import subprocess
 import sys
 import tempfile
 import typing
-from types import MethodType
+import weakref
 from urllib.parse import urlparse
 
 import boto3
@@ -399,6 +400,58 @@ def _infer_credentials(
     )
 
 
+def _temp_file_closer(
+        file: io.IOBase,
+        file_name: str,
+        *upload_args
+):
+    """
+    Close, upload by name, and then delete the file.
+    Meant to replace .close() on a particular instance
+    of a temporary file-like wrapper object, as an unbound
+    callback to a weakref.finalize() registration on the wrapper.
+
+    The reason this implementation is necessary is really complicated.
+
+    ---
+
+    boto3's upload_fileobj could be used before closing the
+    file, instead of closing it and then uploading it by
+    name, but upload_fileobj is less performant than
+    upload_file as of boto3's s3 library s3transfer
+    version 0.6.0.
+
+    For details, see the implementation & comments:
+    https://github.com/boto/s3transfer/blob/0.6.0/s3transfer/upload.py#L351
+
+    TL;DR: s3transfer does multithreaded transfers
+    that require multiple file handles to work properly,
+    but Python cannot duplicate file handles such that
+    they can be accessed in a thread-safe way,
+    so they have to buffer it all in memory.
+    """
+
+    if file.closed:
+        # Makes closure idempotent.
+
+        # If the file object is used as a context
+        # manager, close() is called twice (once in the
+        # serializer code, once after, when leaving the
+        # context).
+
+        # Without this check, this would trigger two
+        # separate uploads.
+        return
+    try:
+        file.close()
+        s3_upload(file_name, *upload_args)
+    finally:
+        try:
+            os.unlink(file_name)
+        except OSError:
+            pass
+
+
 def open_stream(
         path_uri: Union[str, os.PathLike],
         mode: str = "rb",
@@ -468,48 +521,17 @@ def open_stream(
             # with primitive temporary file support (e.g. Windows)
             temp_file = tempfile.NamedTemporaryFile(mode="wb+", delete=False)
 
-            def close(self: temp_file.__class__):
-                # Close, upload by name, and then delete the file.
-                #
-                # boto3's upload_fileobj could be used before closing the
-                # file, instead of closing it and then uploading it by
-                # name, but upload_fileobj is less performant than
-                # upload_file as of boto3's s3 library s3transfer,
-                # version 0.6.0.
-
-                # For details, see the implementation & comments:
-                # https://github.com/boto/s3transfer/blob/0.6.0/s3transfer/upload.py#L351
-
-                # TL;DR: s3transfer does multithreaded transfers
-                # that require multiple file handles to work properly,
-                # but Python cannot duplicate file handles such that
-                # they can be accessed in a thread-safe way,
-                # so they have to buffer it all in memory.
-                if self.closed:
-                    # Makes close() idempotent.
-
-                    # If the resulting object is used as a context
-                    # manager, close() is called twice (once in the
-                    # serializer code, once after, when leaving the
-                    # context).
-
-                    # Without this check, this would trigger two
-                    # separate uploads.
-                    return
-                try:
-                    # Use the original closing method tied to the class,
-                    # rather than the one on this instance. (like super())
-                    self.__class__.close(self)
-                    s3_upload(self.name,
-                              path_uri,
-                              s3_access_key_id,
-                              s3_secret_access_key,
-                              s3_endpoint)
-                finally:
-                    os.unlink(self.name)
-
-            # Bind the method to the instance
-            temp_file.close = MethodType(close, temp_file)
+            guaranteed_closer = weakref.finalize(
+                temp_file,
+                _temp_file_closer,
+                temp_file.file,
+                temp_file.name,
+                path_uri,
+                s3_access_key_id,
+                s3_secret_access_key,
+                s3_endpoint
+            )
+            temp_file.close = guaranteed_closer
             return temp_file
         else:
             s3_endpoint = s3_endpoint or default_s3_read_endpoint