How to hand-invalidate blocks?

[Brian Bockelman <bb...@cse.unl.edu>]

Hey all,

I have a problem with invalid blocks which Hadoop doesn't seem to  
realize that are invalid.

For some reason, a lot of our blocks got truncated to 192KB  
(HADOOP-4543).  When I try to drain off nodes, Hadoop tries to  
replicate these blocks with the truncated blocks as the source.  The  
namenode correctly realizes that the destination copy is incorrectly  
sized, but does not take any action to remove the invalid source  
copy.  Hence, the process just repeats and repeats.

I know the list of the incorrectly sized blocks, and I certainly want  
to fix my cluster before a patch is written, fixed, tested, and  
released.  How do I tell the namenode to forget about the source  
blocks or at least trigger a validation attempt?

If I have a datanode which is suspicious, how do I tell Hadoop to  
verify *all* blocks on that node?  I know validations happen slowly in  
the background, but I don't want to wait weeks before I find out that  
a block has a problem.

Brian