11:51 a.m.
Hi all,
As indicated on IRC, running
org.infinispan.client.hotrod.TopologyChangeTest.testTwoMembers() fails randomly with
replication timeout. It's very easy to replicate. When it fails, this is what
happens:
1. During rehashing, a new hash is installed:
2010-05-06 17:54:11,960 4932 TRACE [org.infinispan.distribution.DistributionManagerImpl]
(Rehasher-eq-985:) Installing new consistent hash DefaultConsistentHash{addresses
={109=eq-35426, 10032=eq-985, 10033=eq-985}, hash space =10240}
2. Rehash finishes and the previous hash is still installed:
2010-05-06 17:54:11,978 4950 INFO [org.infinispan.distribution.JoinTask]
(Rehasher-eq-985:) eq-985 completed join in 30 milliseconds!
3. A put comes in to eq-985 who decides recipients are [eq-985, eq-985]. Most likely, the
hash falled somewhere between 109 and 10032 and since owners are 2, it took the next 2:
2010-05-06 17:54:12,307 5279 TRACE [org.infinispan.remoting.rpc.RpcManagerImpl]
(HotRodServerWorker-2-1:) eq-985 broadcasting call
PutKeyValueCommand{key=CacheKey{data=ByteArray{size=9, hashCode=d28dfa, array=[-84, -19,
0, 5, 116, 0, 2, 107, 48, ..]}}, value=CacheValue{data=ByteArray{size=9, array=[-84, -19,
0, 5, 116, 0, 2, 118, 48, ..]}, version=281483566645249}, putIfAbsent=false,
lifespanMillis=-1000, maxIdleTimeMillis=-1000} to recipient list [eq-985, eq-985]
Everything afterwards is a mess:
4. JGroups removes the local address from the destination. The reason Infinispan does not
do it it's because the number of recipients is 2 and the number of members in the
cluster 2, so it thinks it's a broadcast:
2010-05-06 17:54:12,308 5280 TRACE
[org.infinispan.remoting.transport.jgroups.CommandAwareRpcDispatcher]
(HotRodServerWorker-2-1:) real_dests=[eq-985]
5. JGroups still sends it as a broadcast:
2010-05-06 17:54:12,308 5280 TRACE [org.jgroups.protocols.TCP] (HotRodServerWorker-2-1:)
sending msg to null, src=eq-985, headers are RequestCorrelator: id=201, type=REQ, id=12,
rsp_expected=true, NAKACK: [MSG, seqno=5], TCP: [channel_name=Infinispan-Cluster]
6. Another node deals with this and replies:
2010-05-06 17:54:12,310 5282 TRACE
[org.infinispan.remoting.transport.jgroups.CommandAwareRpcDispatcher]
(OOB-1,Infinispan-Cluster,eq-35426:) Attempting to execute command:
SingleRpcCommand{cacheName='___defaultcache',
command=PutKeyValueCommand{key=CacheKey{data=ByteArray{size=9, hashCode=43487e,
array=[-84, -19, 0, 5, 116, 0, 2, 107, 48, ..]}}, value=CacheValue{data=ByteArray{size=9,
array=[-84, -19, 0, 5, 116, 0, 2, 118, 48, ..]}, version=281483566645249},
putIfAbsent=false, lifespanMillis=-1000, maxIdleTimeMillis=-1000}} [sender=eq-985]
...
7. However, no replies yet from eq-985, so u get:
2010-05-06 17:54:27,310 20282 TRACE
[org.infinispan.remoting.transport.jgroups.CommandAwareRpcDispatcher]
(HotRodServerWorker-2-1:) responses: [sender=eq-985, retval=null, received=false,
suspected=false]
2010-05-06 17:54:27,313 20285 TRACE [org.infinispan.remoting.rpc.RpcManagerImpl]
(HotRodServerWorker-2-1:) replication exception:
org.infinispan.util.concurrent.TimeoutException: Replication timeout for eq-985
Now, I don't understand the reason for creating a hash 10032=eq-985, 10033=eq-985.
Shouldn't keeping 10032=eq-985 be enough? Why add 10033=eq-985?
Assuming there was a valid case for it, a naive approach would be to discard a second node
that points to the an address already in the recipient list. So, 10032=eq-985 would be
accepted for the list but when encountering 10033=eq-985, this would be skipped.
Finally, I thought waiting for rehashing to finish would solve the issue but as u can see
in 2., rehashing finished and the hash is still in the same shape. Also, I've attached
a log file.
Cheers,
--
Galder Zamarreño
Sr. Software Engineer
Infinispan, JBoss Cache