Problem
HadoopUtils keeps global YarnClient/HDFS instances without synchronization. Kerberos ticket refresh closes all clients while watchers and job submission may still be using them, causing intermittent failures.
Proposed solution
Guard YarnClient, HDFS, UGI, and configuration access with a shared lock, including Kerberos refresh timer callbacks.
Problem
HadoopUtilskeeps global YarnClient/HDFS instances without synchronization. Kerberos ticket refresh closes all clients while watchers and job submission may still be using them, causing intermittent failures.Proposed solution
Guard YarnClient, HDFS, UGI, and configuration access with a shared lock, including Kerberos refresh timer callbacks.