Critical Description
There's a bug in FSEditLog#rollEditLog which results in the NN process exiting if a single name dir has failed. Here's the relevant code:
close() // So editStreams.size() is 0 foreach edits dir { .. eStream = new ... // Might get an IOE here editStreams.add(eStream); } catch (IOException ioe) { removeEditsForStorageDir(sd); // exits if editStreams.size() <= 1 }
If we get an IOException before we've added two edits streams to the list we'll exit, eg if there's an error processing the 1st name dir we'll exit even if there are 4 valid name dirs. The fix is to move the checking out of removeEditsForStorageDir (nee processIOError) or modify it so it can be disabled in some cases, eg here where we don't yet know how many streams are valid.
