When brisk accesses cassandra classes inside of a Hadoop Task JVM the AbstractCassandraDaemon uses a log4j PropertyConfigurator to setup cassandra logging. This closes all the existing appenders, including the TaskLogAppender for the hadoop task. They are not opened again because they are not in the config.
log4j has Logger Repositories to handle multiple configs in the same process, but there is a bit of suck involved in making a RepositorySelector.
Basically all the selector has access to thread local storage, and it looks like normally people get the class loader from the current thread. A thread will inherit it's class loader from the thread that created it, unless otherwise specified.
We have code in the same thread the uses hadoop and cassandra classes, so this could be a dead end.
As a work around i've added cassandra.log4j.configure JVM param and made the AbstractCassandraServer skip the log4j config if it's false. My job completes and I can see the cassandra code logging an extra message I put in into the Hadoop task log file...
2011-08-19 15:56:06,442 WARN org.apache.hadoop.metrics2.impl.MetricsSystemImpl: Metrics system not started: Cannot locate configuration: tried hadoop-metrics2-maptask.properties, hadoop-metrics2.properties
2011-08-19 15:56:06,776 INFO org.apache.cassandra.service.AbstractCassandraDaemon: Logging initialized externally
2011-08-19 15:56:07,332 INFO org.apache.hadoop.mapred.MapTask: numReduceTasks: 0
The param has to be passed to the task JVM, so need to modify Haddop mapred-site.xml as follows
Tune your mapred jvm arguments for best performance.
Also see documentation from jvm vendor.
It's not pretty but it works. In my extra log4j logging I can see the second reset() call is gone.
Change the to Hadoop TaskLogAppender also stops the NPE but there may also be some lost log messages