Hello,
We have experienced thread safety issues when subclassing
WritableComparator and subsequently calling WritableComparator.define.
It seems that same comparator instance has its methods called from
different threads. This makes the following code failing from time to
time due to race condition
public class PComparator extends WritableComparator {
....
private DataInputBuffer buf = new DataInputBuffer();
public int compare(byte[] b1, int s1, int l1, byte[] b2, int s2, int
l2) {
... [do something that modifies buf]
}
}
...
static {
WritableComparator.define(PKey.class, new PComparator());
}
So, no member variables can be modified in compare without the explicit
synchronization.
On the other hand, when comparator class is specified in JobConf, each
thread recieives it's own instance and no race condition occurs.
I think this case should be mentioned in documentation
I can also suggest to implement another version
WritableComparator.define(Class c, Class comparatorClass) which
registers the comparaator class, not comparator class instance and make
WritableComparator.get to instantiate comparators registered in such a
way and declare old defile method deprecated
