This is actually due to permission issues in some local yarn catalogs. I started using LinuxContainerExecutor(in unsafe mode with unsafe-mode.local-user as kailash) and made the appropriate changes. However, due to some (unknown) reason, NodeManager was unable to clear local directories for users, and there were still directories with the previous user (in my case, yarn).
, , yarn.nodemanager.local-dirs ( Cloudera, YARN, yarn-site.xml hadoop conf) / node. :
rm -rf /yarn/nm/usercache/*