fix: resolve race condition in FuturesDict and unsafe __del__ methods

[Suraj Sahani (surajsahani)]

Problem

Found and fixed critical concurrency bugs that could cause deadlocks and crashes in production:

1. Race condition in FuturesDict: The callback could fire before the counter was incremented, or a future could complete between counter increment and callback registration, leading to incorrect counter state and potential deadlocks.

2. Unsafe __del__ methods: Multiple stores and cache implementations had __del__ methods that could crash during interpreter shutdown when logging or attributes become unavailable.

Solution

Race Condition Fix (libs/langgraph/langgraph/pregel/_runner.py)

• Increment counter before adding callback to prevent race
• Check if future is already done after callback registration
• Made on_done() idempotent by checking if future was already processed
• Prevents double-decrement of counter

Resource Cleanup Fix

Added proper exception handling in __del__ methods:

• libs/checkpoint-sqlite/langgraph/store/sqlite/base.py
• libs/checkpoint-postgres/langgraph/store/postgres/base.py
• libs/checkpoint-sqlite/langgraph/cache/sqlite/__init__.py
• libs/checkpoint/langgraph/store/base/batch.py

Testing

• ✅ test_pregel_loop_refcount - Memory leak test passes
• ✅ test_concurrent_emit_sends - Concurrency test passes
• ✅ All linting and formatting checks pass

Impact

These fixes prevent:

• Deadlocks in high-concurrency scenarios with many parallel tasks
• Crashes during interpreter shutdown
• Incorrect task counter state leading to hung executions

Particularly important for production workloads with hundreds of concurrent graph executions.