hadoop-hdfs-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From <gu.yiz...@zte.com.cn>
Subject Re: HDFS HA(Based on QJM) Failover Frequently with Large FSimage andBusy Requests
Date Thu, 27 Apr 2017 12:16:30 GMT
1. Is service-rpc configured in namenode?


Not yet, I was considered to configure servicerpc, but I was thinking about the possible disadvantages
as well. 


When failover is happened  because of too many waiting rpcs, if zkfc gets normal process from
another port, is it possiable that the clients get a lot of failures?






2. ha.health-monitor.rpc-timeout.ms - Also consider increasing zkfc rpc call timeout to namenode.



The same worry,  is it possiable that the clients get a lot of failures?






Thanks very much,


Doris










---------------------------------------------------------------------------------------










1. Is service-rpc configured in namenode?
(dfs.namenode.servicerpc-address - this will create another RPC server listening on another
port (say 8021) to handle all service (non-client) requests and hence default rpc address
(say 8020) will handle only client requests.) 

By doing this way, you would be able to decouple client and service requests. Here service
requests corresponds to rpc calls from DN, ZKFC etc. Hence when cluster is too busy because
of too many client operations, ZKFC requests will get processed by different rpc and hence
need not wait in same queue as client requests.)  

2. ha.health-monitor.rpc-timeout.ms - Also consider increasing zkfc rpc call timeout to namenode.


By default this is 45 secs. You can consider increasing it to 1 or 2 mins depending upon your
cluster usage.

Thanks,
Chackra 




On Wed, Apr 26, 2017 at 11:50 AM,  <gu.yizhou@zte.com.cn> wrote:


Hi All,



    HDFS HA (Based on QJM) , 5 journalnodes, Apache 2.5.0 on Redhat 6.5 with JDK1.7.


    Put 1P+ data into HDFS with FSimage about 10G, then keep on making more requests to this
HDFS, namenodes failover frequently. Wanna to know something as follows:






    1.ANN(active namenode) downloading fsimage.ckpt_* from SNN(standby namenode) leads to
very high disk io, at the same time, zkfc fails to monitor the health of ann due to timeout.
Is there any releationship between high disk io and zkfc monitor request timeout? Every failover
happened when ckpt download, but not every ckpt download leads to failover.











2017-03-15 09:27:05,750 WARN org.apache.hadoop.ha.HealthMonitor: Transport-level exception
trying to monitor health of NameNode at nn1/ip:8020: Call From nn1/ip to nn1:8020 failed on
socket timeout exception: java.net.SocketTimeoutException: 45000 millis timeout while waiting
for channel to be ready for read. ch : java.nio.channels.SocketChannel[connected local=/ip:48536
remote=nn1/ip:8020] For more details see:  http://wiki.apache.org/hadoop/SocketTimeout


2017-03-15 09:27:05,750 INFO org.apache.hadoop.ha.HealthMonitor: Entering state SERVICE_NOT_RESPONDING




    2.Due to SERVICE_NOT_RESPONDING, another zkfc fences the old ann(configed sshfence), before
restart by my additional monitor, old ann log sometimes shows like this, what is "Rescan of
postponedMisreplicatedBlocks"? Does this have any reletionships with failover?

2017-03-15 04:36:00,866 INFO org.apache.hadoop.hdfs.server.blockmanagement.CacheReplicationMonitor:
Rescanning after 30000 milliseconds

2017-03-15 04:36:00,931 INFO org.apache.hadoop.hdfs.server.blockmanagement.CacheReplicationMonitor:
Scanned 0 directive(s) and 0 block(s) in 65 millisecond(s).

2017-03-15 04:36:01,127 INFO org.apache.hadoop.hdfs.server.blockmanagement.BlockManager: Rescan
of postponedMisreplicatedBlocks completed in 23 msecs. 247361 blocks are left. 0 blocks are
removed.

2017-03-15 04:36:04,145 INFO org.apache.hadoop.hdfs.server.blockmanagement.BlockManager: Rescan
of postponedMisreplicatedBlocks completed in 17 msecs. 247361 blocks are left. 0 blocks are
removed.

2017-03-15 04:36:07,159 INFO org.apache.hadoop.hdfs.server.blockmanagement.BlockManager: Rescan
of postponedMisreplicatedBlocks completed in 14 msecs. 247361 blocks are left. 0 blocks are
removed.

2017-03-15 04:36:10,173 INFO org.apache.hadoop.hdfs.server.blockmanagement.BlockManager: Rescan
of postponedMisreplicatedBlocks completed in 14 msecs. 247361 blocks are left. 0 blocks are
removed.

2017-03-15 04:36:13,188 INFO org.apache.hadoop.hdfs.server.blockmanagement.BlockManager: Rescan
of postponedMisreplicatedBlocks completed in 14 msecs. 247361 blocks are left. 0 blocks are
removed.

2017-03-15 04:36:16,211 INFO org.apache.hadoop.hdfs.server.blockmanagement.BlockManager: Rescan
of postponedMisreplicatedBlocks completed in 23 msecs. 247361 blocks are left. 0 blocks are
removed.

2017-03-15 04:36:19,234 INFO org.apache.hadoop.hdfs.server.blockmanagement.BlockManager: Rescan
of postponedMisreplicatedBlocks completed in 22 msecs. 247361 blocks are left. 0 blocks are
removed.

2017-03-15 04:36:28,994 INFO org.apache.hadoop.hdfs.server.namenode.NameNode: STARTUP_MSG:






    3.I config two dfs.namenode.name.dir and one dfs.journalnode.edits.dir(which shares one
disk with nn), is it suitable? Or does this have any disadvantage?








<property>

<name>dfs.namenode.name.dir.nameservice.nn1</name>

<value>/data1/hdfs/dfs/name,/data2/hdfs/dfs/name</value>

</property>

<property>

<name>dfs.namenode.name.dir.nameservice.nn2</name>

<value>/data1/hdfs/dfs/name,/data2/hdfs/dfs/name</value>

</property>








<property>

<name>dfs.journalnode.edits.dir</name>

<value>/data1/hdfs/dfs/journal</value>

</property>


    



    4.Interested in design of checkpoint and edit logs transmission,any explanation,issues
or documents?







Thanks in advance,

Doris
Mime
View raw message