Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
Description
map = Collections.synchronizedMap( new WeakHashMap<String, WeakReference<Class<?>>>());
This synchronizes all lookups across the same class-loader across all threads & yields rpc threads.
When reading from HDFS with good locality, this fills up the contended lock profile with almost no other contributors to the locking - see llap-rpc-locks.svg