Both lookup_coordinator and _send_heartbeat_request make client api calls that require locking. The tricky part of the heartbeat run loop is that we need to always acquire client lock first and coordinator lock second (to avoid deadlocks), but we also cannot hold the client lock while calling coordinator lock.wait() . To make this easier, switch from context managers to try / finally and catch any RuntimeError if the client lock has already been released when we reach the finally block.