Details
Description
We ran into the following sequence of events:
- master startup failed after only ROOT had been assigned (for another reason)
- restarted the master without restarting other servers. Since there was at least one region assigned, it went through the failover code path
- master scanned META and inserted every region into /hbase/unassigned in ZK.
- then, it called "listChildren" on the /hbase/unassigned znode, and crashed with "Packet len6080218 is out of range!" since the IPC response was larger than the default maximum.
Issue Links
- is broken by
-
ZOOKEEPER-1162
consistent handling of jute.maxbuffer when attempting to read large zk "directories"
-
- Open
-
- is related to
-
HBASE-6625
If we have hundreds of thousands of regions getChildren will encouter zk exception
-
- Resolved
-
Activity
Moving out of 0.94. Pull back if you feel otherwise.
Hi, It also happpened in our cluster when we restarted whole cluster(it has 129723 regions).
2012-06-19 19:29:00,961 DEBUG org.apache.hadoop.hbase.zookeeper.ZKAssign: master:20000-0x137ed2eb936fb85 Creating (or updating) unassigned node for 80400ccd4a1f3438cc23774ca8a88d17 with OFFLINE state
2012-06-19 19:29:00,965 DEBUG org.apache.hadoop.hbase.master.AssignmentManager: Handling transition=M_ZK_REGION_OFFLINE, server=172-16-6-2:20000, region=80400ccd4a1f3438cc23774ca8a88d17
2012-06-19 19:29:00,966 DEBUG org.apache.hadoop.hbase.zookeeper.ZKAssign: master:20000-0x137ed2eb936fb85 Creating (or updating) unassigned node for 7f1a56641906ae0a6cc6919bd927df76 with OFFLINE state
2012-06-19 19:29:00,969 DEBUG org.apache.hadoop.hbase.master.AssignmentManager: Handling transition=M_ZK_REGION_OFFLINE, server=172-16-6-2:20000, region=7f1a56641906ae0a6cc6919bd927df76
2012-06-19 19:29:01,070 WARN org.apache.zookeeper.ClientCnxn: Session 0x137ed2eb936fb85 for server 172-16-6-1/172.16.6.1:2181, unexpected error, closing socket connection and attempting reconnect
2012-06-19 19:29:01,070 WARN org.apache.zookeeper.ClientCnxn: Session 0x137ed2eb936fb85 for server 172-16-6-1/172.16.6.1:2181, unexpected error, closing socket connection and attempting reconnect
java.io.IOException: Packet len4670048 is out of range!
at org.apache.zookeeper.ClientCnxn$SendThread.readLength(ClientCnxn.java:721)
at org.apache.zookeeper.ClientCnxn$SendThread.doIO(ClientCnxn.java:880)
at org.apache.zookeeper.ClientCnxn$SendThread.run(ClientCnxn.java:1145)
2012-06-19 19:29:01,174 WARN org.apache.hadoop.hbase.zookeeper.ZKUtil: master:20000-0x137ed2eb936fb85 Unable to list children of znode /hbase/unassigned
@Gao
You got this in 0.90? or someother version?
This may come in latest version also as we didn't change the znode hierarchy of the unassigned regions. As mentioned in linked issue, there is a cap on packet length. We can't read/write huge data in a single packet.
IMO, to resolve this we need to do either of the following.
- In HBASE: We can use hierarchical structure.
HDFS datanode follows similar strategy. It keeps block files in different sub directories to avoid FS lookup latency.
- In ZooKeeper: Increase the limit. What is reasonable?
We have tried this out in some other project but it has the side effects. When we tried read/write huge data from ZooKeeper, clients occassionally gets disconnected. This is sequential request processing. Please check out the related discussions @
Following JIRA and discussion also applicable in current scenario.
http://mail-archives.apache.org/mod_mbox/zookeeper-user/201104.mbox/%3CFFA3BDB6-1C83-42B9-B2A0-7675134626C5@me.com%3E
https://issues.apache.org/jira/browse/ZOOKEEPER-1049
The version is 0.90.X, I have asked the customer up jute.maxbuffer to 64M.
Am downgrading the priority since there is a work-around (increase the buffer size).
I added a note to troubleshooting doc in the master startup section.
Integrated in HBase-TRUNK #3759 (See https://builds.apache.org/job/HBase-TRUNK/3759/)
Add to troubleshooting a note on zk buffer size issue when lots of regions – hbase-4246 (Revision 1434087)
Result = FAILURE
Integrated in HBase-TRUNK-on-Hadoop-2.0.0 #351 (See https://builds.apache.org/job/HBase-TRUNK-on-Hadoop-2.0.0/351/)
Add to troubleshooting a note on zk buffer size issue when lots of regions – hbase-4246 (Revision 1434087)
Result = FAILURE
Not being worked on for 0.95. Moving it out. Leaving it around in case others run into the issue so they can find the workaround. We should also consider doing the suggested hierarcherying (word?) in zk).
We were able to work around this issue by bumping jute.maxbuffer up to 100MB on the cluster in question.
Another solution would be to shard the /hbase/unassigned dir by a prefix of the region ID. eg region 1234567890abcdef would go in /hbase/unassigned/1234/1234567890abcdef - so, we have to do a traversal to get the full list, but any particular RPC response is limited in size.