Skip to content

fix(batch): release pooled model slots on failures and cancellation - #763

Open
YMuskrat wants to merge 1 commit into
NVIDIA:mainfrom
YMuskrat:ymuskrat/pool-slot-cleanup
Open

YMuskrat wants to merge 1 commit into
NVIDIA:mainfrom
YMuskrat:ymuskrat/pool-slot-cleanup

Conversation

@YMuskrat

@YMuskrat YMuskrat commented Oct 5, 2026

Copy link
Copy Markdown
Contributor

Fixes #760.

PooledChatModel reserves a slot before constructing ChatOpenAI, but construction currently happens outside cleanup. Repeated construction errors can exhaust the pool even though no request is running. Cancelled async calls also skip cleanup, and cancellation while waiting on a threaded blocking acquire can leave an orphaned waiter that reserves capacity later.

Put client construction and invocation inside a try/finally release in both retry loops. For async capacity waiting, retry the existing non-blocking acquire after a cancellable 50 ms sleep, avoiding a background acquiring thread. This adds up to one polling interval of capacity-wakeup latency when the pool is full. Existing rate-limit backoff, retries and callback forwarding are retained.

This is per-call cleanup, complementary to the owner/process cleanup proposed in #743.

Validation: make lint, make format, and make format-check pass; all 9,965 unit tests and 124 integration tests pass. Seven new cases cover construction failures, in-flight cancellation, waiting cancellation, capacity recovery and sync/async rate-limit retries. Four fail on the original implementation. The focused run including the existing contrib pool tests passes all 36 tests; no live provider calls were made.

Release reserved slots in finally blocks and keep asynchronous capacity waiting cancellable.

Signed-off-by: YMuskrat <101849520+YMuskrat@users.noreply.github.com>

@yashrajp22 yashrajp22 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I found no actionable introduced defect in the scoped review of head 6df3025 against base b6f0304. All 65 focused HEAD pool checks passed in both dependency modes, covering construction/invocation failures and cancellation. The pool code was tested from contrib source in both modes; it is not shipped in the wheel.

The runtime evidence remains qualified: the harness continued after a supplemental preflight failure, so the recorded launch-sequencing deviation still applies. Provider calls were mocked. Combined callbacks-on-retry, live cancellation/transport cleanup, load performance and fairness, other Python/platform combinations, and the full repository suite were not verified. Nine of the 12 corpus samples per mode remained partial, with signature exclusions retained; this is not a global accuracy result. This summary applies to the pinned comparison above; integration with newer main has not been reviewed.

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.

PooledChatModel leaks API-key slots on client construction failures and cancellation

2 participants