cloudpipe/cloudpickle

Class tracking in the main process silences potential dynamic classes test failures

Open

#313 opened on 2019年10月26日

GitHub で見る
 (0 comments) (1 reaction) (0 assignees)Python (1,407 stars) (152 forks)batch import
help wanted

説明

Now that cloudpickle tracks dynamically created classes, pickle_depickle(some_dynamic_class) does not creates a new class anymore, it just returns the one already present in the main process.

As an example if we change the reproducer of #311 by using cloudpickle.loads(cloudpickle.dumps(a)) instead of dumping a pickle in a file and loading it in a fresh python interpreter, we don't get the reported error.

Would we want to be rigorous, we would probably have to test everything involving dynamic class construction in both settings (untracked class depickled for the first time in a child process, tracked class depickled from either a warm child process or the parent process).

コントリビューターガイド

技術スタック
python
領域
testing
Issue 種別
test
難度新しい貢献者にとっての実装難度です。1 は非常に小さな変更、5 は専門的な作業です。
3
推定時間経験ある貢献者が調査、実装、テスト、pull request 準備にかけるおおよその時間範囲です。
half day
活動状況Issue が今どれだけ取り組みやすいかを示します。新しい、活発、古い、ブロック中、またはメンテナー入力待ちなどです。
stale
明確さ期待される変更、受け入れ条件、次の手順がどれだけ明確かを示します。
mostly clear
前提条件
Pythoncloudpicklebasic understanding of pickling
初心者向け度初回貢献者にどれだけ取り組みやすいかを 1-100 で推定したスコアです。
30
調査方針
Investigate the cloudpickle source code, focusing on the dynamic class tracking mechanism in `cloudpickle.cloudpickle.py`. Review the original issue #311 and its reproducer to understand the expected behavior. Write a test that runs in both a parent process and a child process to verify that dynamic class pickling works correctly in both settings. Consider adding a cross process test using `multiprocessing` or `subprocess` to ensure no regression.