Hi All,

We are currently testing our app for performance and found a strange effect.

We load large amount of data using Hector 1.0-5 + Cassandra Thrift 1.0.9.

Total amount of distinct keys is 3M. Each key(K) has one composite column(C) and binary array as value(V).

We write it using mutator.addInsertion(...) / execute().

After some execution time you can see that HConnectionManager holds 800M of byte[] data which is not eligible for GC.

Setting -Xmx1500m on client JVM helps to solve the problem but we would like to run our app with much smaller memory footprint.

Can anyone please give any advice to fight the memory issue?