Slow shutdown due to initial replication

[nojima]

The shutdown process of yrmcds waits for the gc_thread to finish after stopping the reactor. Sometimes, the gc_thread is performing initial replication, which can take a long time.

As soon as the master’s reactor stops, the slave detects the master’s failure and tries to become the new master. However, the master process is not completely terminated until the gc_thread finishes, so the VIP does not move to the slave. If the slave cannot become the master within 5 seconds, it discards its own data. This leads to data loss in the cluster.

I am not sure what the best solution is, but I believe some kind of countermeasure is necessary.