The RegionServer hosting hbase:meta will certainly have "load" placed
onto it, commensurate to the size of your cluster and the number of
clients you're running. However, this shouldn't be increasing the amount
of connections to ZK from a RegionServer.
The RegionServer hosting system.catalog would be unique WRT other
Phoenix-table Regions. I don't recall off of the top of my head if there
is anything specific in the RegionServer code that runs alongside
system.catalog (the MetaDataEndpoint protocol) that reaches out to
ZooKeeper.
If you're using HDP 2.6.3, I wouldn't be surprised if you're running
into known and fixed issues where ZooKeeper connections are not cleaned
up. That's multiple-years old code.
netstat and tcpdump isn't really going to tell you anything you don't
already. From a thread dump or a heap dump, you'll be able to see the
number of ZooKeeper connections from a RegionServer. The 4LW commands
from ZK will be able to tell you which clients (i.e. RegionServers) have
the most connections. These numbers should match (X connections from a
RS to a ZK, and X connections in the Java RS process). The focus would
need to be on what opens a new connection and what is not properly
closing that connection (in every case).
On 6/3/20 4:57 AM, anil gupta wrote:
Thanks for sharing insights. Moving hbase mailing list to cc.
Sorry, forgot to mention that we are using Phoenix4.7(HDP 2.6.3). This
cluster is mostly being queried via Phoenix apart from few pure NoSql
cases that uses raw HBase api's.
I looked further into zk logs and found that only 6/15 RS are running
into max connection problems(no other ip/hosts of our client apps were
found) constantly. One of those RS is getting 3-4x the connections
errors as compared to others, this RS is hosting hbase:meta
<http://ip-10-74-10-228.us-west-2.compute.internal:16030/region.jsp?name=1588230740>,
regions of phoenix secondary indexes and region of Phoenix and HBase
tables. I also looked into other 5 RS that are getting max connection
errors, for me nothing really stands out since all of them are hosting
regions of phoenix secondary indexes and region of Phoenix and HBase tables.
I also tried to run netstat and tcpdump on zk host to find out anomaly
but couldn't find anything apart from above mentioned analysis. Also ran
hbck and it reported that things are fine. I am still unable to pin
point exact problem(maybe something with phoenix secondary index?). Any
other pointer to further debug the problem will be appreciated.
Lastly, I constantly see following zk connection loss logs in above
mentioned 6 RS:
/2020-06-03 06:40:30,859 WARN
[RpcServer.FifoWFPBQ.default.handler=123,queue=3,port=16020-SendThread(ip-10-74-0-120.us-west-2.compute.internal:2181)] zookeeper.ClientCnxn: Session 0x0 for server ip-10-74-0-120.us-west-2.compute.internal/10.74.0.120:2181 <http://10.74.0.120:2181>, unexpected error, closing socket connection and attempting reconnect
java.io.IOException: Connection reset by peer
at sun.nio.ch.FileDispatcherImpl.read0(Native Method)
at sun.nio.ch.SocketDispatcher.read(SocketDispatcher.java:39)
at sun.nio.ch.IOUtil.readIntoNativeBuffer(IOUtil.java:223)
at sun.nio.ch.IOUtil.read(IOUtil.java:192)
at sun.nio.ch.SocketChannelImpl.read(SocketChannelImpl.java:380)
at
org.apache.zookeeper.ClientCnxnSocketNIO.doIO(ClientCnxnSocketNIO.java:68)
at
org.apache.zookeeper.ClientCnxnSocketNIO.doTransport(ClientCnxnSocketNIO.java:366)
at
org.apache.zookeeper.ClientCnxn$SendThread.run(ClientCnxn.java:1125)
2020-06-03 06:40:30,861 INFO
[RpcServer.FifoWFPBQ.default.handler=137,queue=17,port=16020-SendThread(ip-10-74-9-182.us-west-2.compute.internal:2181)] zookeeper.ClientCnxn: Opening socket connection to server ip-10-74-9-182.us-west-2.compute.internal/10.74.9.182:2181 <http://10.74.9.182:2181>. Will not attempt to authenticate using SASL (unknown error)
2020-06-03 06:40:30,861 INFO
[RpcServer.FifoWFPBQ.default.handler=137,queue=17,port=16020-SendThread(ip-10-74-9-182.us-west-2.compute.internal:2181)] zookeeper.ClientCnxn: Socket connection established, initiating session, client: /10.74.10.228:60012 <http://10.74.10.228:60012>, server: ip-10-74-9-182.us-west-2.compute.internal/10.74.9.182:2181 <http://10.74.9.182:2181>
2020-06-03 06:40:30,861 WARN
[RpcServer.FifoWFPBQ.default.handler=137,queue=17,port=16020-SendThread(ip-10-74-9-182.us-west-2.compute.internal:2181)] zookeeper.ClientCnxn: Session 0x0 for server ip-10-74-9-182.us-west-2.compute.internal/10.74.9.182:2181 <http://10.74.9.182:2181>, unexpected error, closing socket connection and attempting reconnect
java.io.IOException: Connection reset by peer
at sun.nio.ch.FileDispatcherImpl.read0(Native Method)
at sun.nio.ch.SocketDispatcher.read(SocketDispatcher.java:39)
at sun.nio.ch.IOUtil.readIntoNativeBuffer(IOUtil.java:223)
at sun.nio.ch.IOUtil.read(IOUtil.java:192)
at sun.nio.ch.SocketChannelImpl.read(SocketChannelImpl.java:380)
at
org.apache.zookeeper.ClientCnxnSocketNIO.doIO(ClientCnxnSocketNIO.java:68)
at
org.apache.zookeeper.ClientCnxnSocketNIO.doTransport(ClientCnxnSocketNIO.java:366)
at
org.apache.zookeeper.ClientCnxn$SendThread.run(ClientCnxn.java:1125)/
Thanks!
On Tue, Jun 2, 2020 at 6:57 AM Josh Elser <els...@apache.org
<mailto:els...@apache.org>> wrote:
HBase (daemons) try to use a single connection for themselves. A RS
also
does not need to mutate state in ZK to handle things like gets and puts.
Phoenix is probably the thing you need to look at more closely
(especially if you're using an old version of Phoenix that matches the
old HBase 1.1 version). Internally, Phoenix acts like an HBase client
which results in a new ZK connection. There have certainly been bugs
like that in the past (speaking generally, not specifically).
On 6/1/20 5:59 PM, anil gupta wrote:
> Hi Folks,
>
> We are running in HBase problems due to hitting the limit of ZK
> connections. This cluster is running HBase 1.1.x and ZK 3.4.6.x
on I3en ec2
> instance type in AWS. Almost all our Region server are listed in
zk logs
> with "Too many connections from /<IP> - max is 60".
> 2020-06-01 21:42:08,375 - WARN [NIOServerCxn.Factory:
> 0.0.0.0/0.0.0.0:2181:NIOServerCnxnFactory@193
<http://0.0.0.0/0.0.0.0:2181:NIOServerCnxnFactory@193>] - Too many
connections from
> /<ip> - max is 60
>
> On a average each RegionServer has ~250 regions. We are also
running
> Phoenix on this cluster. Most of the queries are short range
scans but
> sometimes we are doing full table scans too.
>
> It seems like one of the simple fix is to increase maxClientCnxns
> property in zoo.cfg to 300, 500, 700, etc. I will probably do
that. But, i
> am just curious to know In what scenarios these connections are
> created/used(Scans/Puts/Delete or during other RegionServer
operations)?
> Are these also created by hbase clients/apps(my guess is NO)? How
can i
> calculate optimal value of maxClientCnxns for my cluster/usage?
>
--
Thanks & Regards,
Anil Gupta
