[Sarbajit-Sengupta]

Motivation and context

The in-memory fallback cache checks membership and then indexes the same dictionaries. A concurrent eviction, clear or invalidation can remove a user or flag between those operations, raising KeyError. The stale read is used from the public flag lookup's request-error handlers, so a cache eviction during an outage can itself escape to application code.

Read each dictionary entry with get() and keep that result locally. An entry removed before its lookup is an ordinary miss. This retains generation, TTL, stale-age and result checks without adding a blocking cache lock or changing public API. It is a narrow lookup-race fix, not a claim that every cache operation is now thread-safe.

How tested

• Six deterministic regression cases fail on unchanged upstream and pass with the fix: user eviction and flag removal on fresh/stale reads, plus public get_feature_flag_result() calls during a simulated connection outage.
• A separate native-dictionary threading probe with forced yields between reader statements produced 714 KeyErrors before / zero after (four readers, 30,000 iterations each). The injected yields deliberately exercise the race; these counts are not a production failure-rate claim.
• Utility, cache-publication, Redis-snapshot and new regression suites: 179 passed, 2 skipped on Windows / Python 3.13, with loopback access enabled.
• Whole-repository Ruff formatting and lint checks passed; the prescribed mypy/baseline pipeline passed for 326 source files; warning-as-error import and diff checks passed.
• The broader SDK run reached 2,750 passed, 26 skipped, 18 subtests passed, then stopped at TestClient.test_basic_capture_exception_with_no_exception_given, whose filename assertion expects / while Windows emits \\. The same test fails on unchanged upstream code. A complete green full-suite run is not claimed.

Checklist

• All commits are signed (GitHub reports a valid signature).
• I reviewed the submitted code. (Automated review performed; independent human review is not claimed.)
• I added tests to verify the changes.
• No public API signature or documentation change is needed.

Agent context

Autonomy: Human-driven (agent-assisted). The account owner requested tested open-source contributions; the agent selected this defect, implemented it and performed the validation above. The human DRI is @Sarbajit-Sengupta. Assignment to that account was requested when creating the PR, but GitHub denied ReplaceActorsForAssignable; a maintainer needs to apply the assignee.

OpenAI Codex desktop, Python/pytest, Ruff, mypy, Git and GitHub CLI were used. The model is a GPT-6-based coding agent; an exact deployment variant is unavailable. A public session link is omitted because the conversation includes private account/workspace context. This description is not a claim of human production experience, independent human code review, a live-provider test or guaranteed delivery during outages. Human maintainer review is required; no self-merge is requested.

[marandaneto]

The code looks good; no qualifying findings. Verdict: correct.

[marandaneto]

thanks @Sarbajit-Sengupta