Any other ideas?Sounds like a nasty heisenbug, can you replace or rebuild the machine?
I've finally had some time to experiment a bit with this problem (it occured twice again) and here's what I found:
1. So far (three occurences in total), *when* it happened, it happened only for streaming to *one* specific C* node (but it works on this node too for 99,9% of the time)
2. It happens with compression turned on (cassandra.output.compression.class set to org.apache.cassandra.io.compress.DeflateCompressor, but it doesn't matter what the chunk length is)
3. Everything works fine when compression is turned off.
So it looks like I have a workaround for now, but I don't really understand the root cause of this problem and what's the "right" solution if we want to keep using compression.
Anyway, the thing that interests me the most is why does it fail so rarely and - assuming it's not a coincidence - why only for one C* node...
May it be a DeflateCompressor's bug?
Any other ideas?
W dniu 31.03.2013 12:01, aaron morton pisze:but yesterday one of 600 mappers failed:)From what I can understand by looking into the C* source, it seems to me that the problem is caused by a empty (or surprisingly finished?) input buffer (?) causing token to be set to -1 which is improper for RandomPartitioner:Yes, there is a zero length key which as a -1 token.However, I can't figure out what's the root cause of this problem.mmm, the BulkOutputFormat uses a SSTableSimpleUnsortedWriter and neither of them check for zero length row keys. I would look there first.
There is no validation in the AbstractSSTableSimpleWriter, not sure if that is by design or an oversight. Can you catch the zero length key in your map job ?
Freelance Cassandra Consultant
On 28/03/2013, at 2:26 PM, Michal Michalski <firstname.lastname@example.org> wrote:We're streaming data to Cassandra directly from MapReduce job using BulkOutputFormat. It's been working for more than a year without any problems, but yesterday one of 600 mappers faild and we got a strange-looking exception on one of the C* nodes.
IMPORTANT: It happens on one node and on one cluster only. We've loaded the same data to test cluster and it worked.
ERROR [Thread-1340977] 2013-03-28 06:35:47,695 CassandraDaemon.java (line 133) Exception in thread Thread[Thread-1340977,5,main]
java.lang.RuntimeException: Last written key DecoratedKey(5664330507961197044404922676062547179, 302c6461696c792c32303133303332352c312c646f6d61696e2c756e6971756575736572732c633a494e2c433a6d63635f6d6e635f636172726965725f43656c6c4f6e655f4b61726e6174616b615f2842616e67616c6f7265295f494e2c643a53616d73756e675f47542d49393037302c703a612c673a3133) >= current key DecoratedKey(-1, ) writing into /cassandra/production/IndexedValues/production-IndexedValues-tmp-ib-240346-Data.db
From what I can understand by looking into the C* source, it seems to me that the problem is caused by a empty (or surprisingly finished?) input buffer (?) causing token to be set to -1 which is improper for RandomPartitioner:
public BigIntegerToken getToken(ByteBuffer key)
if (key.remaining() == 0)
return MINIMUM; // Which is -1
return new BigIntegerToken(FBUtilities.hashToBigInteger(key));
However, I can't figure out what's the root cause of this problem.
Of course I can't exclude a bug in my code which streams these data, but - as I said - it works when loading the same data to test cluster (which has different number of nodes, thus different token assignment, which might be a case too).