Skip to content

repro: add tests reproducing Session data races under free-threading#3727

Draft
crowecawcaw wants to merge 1 commit into
boto:developfrom
crowecawcaw:nogil-race-repro
Draft

repro: add tests reproducing Session data races under free-threading#3727
crowecawcaw wants to merge 1 commit into
boto:developfrom
crowecawcaw:nogil-race-repro

Conversation

@crowecawcaw

Copy link
Copy Markdown

A single botocore.session.Session lazily builds and caches shared state (components, credentials, parsed config) and HierarchicalEmitter mutates a shared handler trie, using check-then-act without synchronization. These races are latent under the GIL (the ~5ms switch interval makes the window rare) but reproduce when a Session is shared across threads on a free-threaded (no-GIL) build, where threads run Python in parallel.

These tests reproduce four such races and are written to fail on the current code.

They use delay-injection (a sleep in the lazy-init window, so the factory should still run exactly once) and identity invariants (all threads should observe the same singleton), and shrink the thread-switch interval to widen the window on the GIL build.

Tests generated with AI assistance and reviewed by me. Not exactly intended to be merged as is, but more to demonstrate the issue.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

A single botocore.session.Session lazily builds and caches shared state
(components, credentials, parsed config) and HierarchicalEmitter mutates
a shared handler trie, using check-then-act without synchronization.
These races are latent under the GIL (the ~5ms switch interval makes the
window rare) but reproduce when a Session is shared across threads on a
free-threaded (no-GIL) build, where threads run Python in parallel.

These tests reproduce four such races and are written to fail on the
current code:

- ComponentLocator.get_component builds a deferred component more than
  once under concurrent first-access.
- Session.get_credentials loads credentials more than once.
- Session.full_config parses the config files more than once.
- HierarchicalEmitter concurrent registration loses handlers via a
  check-then-act node creation in _PrefixTrie.append_item.

They use delay-injection (a sleep in the lazy-init window, so the factory
should still run exactly once) and identity invariants (all threads
should observe the same singleton), and shrink the thread-switch interval
to widen the window on the GIL build. Each fails deterministically on
both the default and free-threaded builds, and passes once the shared
state is synchronized.

Tests generated with AI assistance and reviewed by me.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant