Context

PYTHON-5672 eliminates three generator frames from the checkout path. The next bottleneck is lock granularity.

Topology lock: topology.select_server() holds a single coarse lock for the entire selection loop, serializing every concurrent checkout even when a suitable server is already cached. Replace it with a reader/writer pattern: reads of topology.description hold only a shared read lock; the write lock is acquired only during monitor updates (topology mutations). This is the same approach used in the Java and Go drivers.

Pool lock merging: _get_conn() }}acquires two back-to-back locks on the fast path ({{operation_count then active_sockets). Merge into one acquire to remove ~50–150 ns per checkout.

Definition of done

Locks are optimized, CMAP, SDAM, and unified-format spec tests pass, concurrent throughput under 10+ parallel checkouts improves measurably.

Pitfalls

Ensure we are not introducing undo complexity.