Fsck shows one of the files in our dfs cluster is corrupt.
/bin/hadoop fsck aFile -files -blocks -locations
aFile: 4633 bytes, 2 block(s):
aFile: CORRUPT block blk_-4597378336099313975
0. blk_-4597378336099313975_2284630101 len=0 repl=3 [...]
1. blk_5024052590403223424_2284630107 len=4633 repl=3 [...]Status: CORRUPT
On disk, these two blocks are of the same size and the same content. It turns out the writer of the file is from a multiple threaded map task. Each thread may write to the same file. One possible interaction of two threads might make this to happen:
[T1: create aFile] [T2: delete aFile] [T2: create aFile][T1: addBlock 0 to aFile][T2: addBlock1 to aFile]...
Because T1 and T2 have the same client name, which is the map task id, the above interactions could be done without any lease exception, thus eventually leading to a corrupt file. To solve the problem, a mapreduce task's client name could be formed by its task id followed by a random number.