Affects Version/s: 0.5.0
Fix Version/s: 0.10.0
If the target path of the dfs -mv command exists within the source path, the dfs becomes corrupt. For example:
% hadoop dfs -mkdir target
% hadoop dfs -mv / target
I'm not certain whether this is reproducible in the current trunk, but I'd bet that it is.
This problem successfully circumvented my own patch to make dfs -rm a little safer (see my email c.2006-08-30 to nutch-dev for details). I had been deleting old crawl directories from the DFS by copying their names and pasting them into my command buffer. At one point, I paused to do something else, copied some other text (which unfortunately began with a Java comment and included carriage returns), then went back to removing the crawl directories. I must not have pressed hard enough on the "c" key when I did my next copy, since when I pasted into the command buffer, hadoop immediately began executing a dfs -rm / command. No problem - I'm protected, because my patched dfs command is just going to try to move / to /trash (and fail), right?
Wrong! Even though hadoop isn't really capable of such a move, it apparently tries hard enough to corrupt the namenode's DB.
Thankfully, I ran into this problem at a relatively opportune time, when the contents of my dfs had little value.