It turns out that in 2.0 we changed the way full snapshots are sent from Sentry to HDFS. Before they were using HMSPaths which used tree structure and eliminated some duplication. Also
SENTRY-1827 helped to compressed this on the serialization side.
Now we are using TPathChanges structure that is not tree-based and contains very non-efficient way of representing paths: required list<list<string>> addPaths; so we split each paths on slashes and store list of elements instead of storing a tree. As a result we may use much more memory.