DDevelopment & Design

[:en] Operational problems with Zookeeper

This post is a summary of what has been presented by Kathleen Ting on StrangeLoop conference. You can watch the original here: http://www.infoq.com/presentations/Misconfiguration-ZooKeeper I’ve decided to put this selection here for quick reference. …This post is a summary of what has been presented by Kathleen Ting on StrangeLoop conference. You can watch the original here: http://www.infoq.com/presentations/Misconfiguration-ZooKeeper I’ve decided to put this selection here for quick reference. …

This post is a summary of what has been presented by Kathleen Ting on
StrangeLoop conference. You can watch the original here:
http://www.infoq.com/presentations/Misconfiguration-ZooKeeper

I’ve decided to put this selection here for quick reference.

Connection mismanagement

  • too many connections 
    WARN [NIOServerCxn.Factory: 0.0.0.0/0.0.0.0:2181:NIOServerCnxn$Factory@247] - Too many connections from /xx.x.xx.xxx - max is 60
  • running out of ZK connections?
    • set maxClientCnxns=200 in zoo.cfg
  • HBase client leaking connections?
    • fixed in HBASE-3777, HBASE-4773, HBASE-5466
    • manually close connections
  • connection closes prematurely 
    ERROR: org.apache.hadoop.hbase.ZooKeeperConnectionException: HBase is able to connect to ZooKeeper but the connection closes immediately.
  • in hbase-site.xml set hbase.zookeeper.recoverable.waittime=30000ms
  • pig hangs connecting to HBase

    CAUSE: location of ZK quorum is not known to Pig

    • use Pig 10, which includes PIG-2115
    • if there is an overlap between TaskTrackers and ZK quorum nodes
      • set hbase.zookeeper.quorum to final in hbase-site.xml
      • otherwise add hbaze.zoopeeker.quorum=hadoophbasemaster.lan:2181 in pig.properties
WARN org.apache.zookeeper.ClientCnxn: Session 0x0 for server null, unexpected error, closing socket connection and attempting reconnect java.net.ConnectionException: Connection refused!

Time mismanagement

  • client session timed out
INFO org.apache.zookeeper.server.ZooKeeperServer: Expiring session <id>, timeout of 40000ms exceeded
    • ZK and HBase need the same session timeout values
      • zoo.cfg: maxSession=Timeout=180000
      • hbase-site.xml: zookeeper.session.timeout=180000
    • don’t co-locate ZK with IO-intense DataNode or RegionServer
    • specify right amount of heap and tune GC flags
      • turn on parallel/CMS/incremental GC
  •  
  • clients lose connections 
    WARN org.apache.zookeeper.ClientCnxn - Session <id> for server <name>, unexpected error, closing socket connection and attempting reconnect java.io.IOException: Broken pipe
    • don’t use SSD drive for ZK transaction log

Disk management

  • unable to load database – unable to run quorum server 
    FATAL Unable to load database on disk !  java.io.IOException: Failed to process transaction type: 2 error: KeeperErrorCode = NoNode for <file> at org.apache.zookeeper.server.persistence.FileTxnSnapLog.restore(FileTxnSnapLog.java:152)!
    • archive and wipe /var/zookeeper/version-2 if other two ZK servers
      are running
  • unable to load database – unreasonable length exception 
    FATAL Unable to load database on disk java.io.IOException: Unreasonable length = 1048583 at org.apache.jute.BinaryInputArchive.readBuffer(BinaryInputArchive.java:100)
    • server allows a client to set data larger than the server can read from disk
    • if a znode is not readable, increase jute.maxbuffer
      • look for "Packet len <xx> is out of range" in the client log
      • increase it by 20%
      • set in JVMFLAGS="-Djute.maxbuffer=yy" bin/zkCli.sh
      • fixed in ZOOKEEPER-151
  • failure to follow leader 
    WARN org.apache.zookeeper.server.quorum.Learner: Exception when following the leader java.net.SocketTimeoutException: Read timed out

    CAUSE:

    • disk IO contention, network issues
    • ZK snapshot is too large (lots of ZK nodes)

    SOLVE:

    • reduce IO contention by putting dataDir on dedicated spindle
    • increase initLimit on all ZK servers and restart, see
      ZOOKEEPER-1521
    • monitor network

Best Practices

DOs

  • separate spindles for dataDir & dataLogDir
  • allocate 3 or 5 ZK servers
  • tune garbage collection
  • run zkCleanup.sh script via cron

DON’Ts

  • dont’ co-locate ZK with I/O intense DataNode or RegionServer
  • don’t use SSD drive for ZK transaction log

You may use Zookeeper as an observer – a non-voting member:

  • in zoo.cfg 
    peerType=observer

— Previous article

TouK na warsztatach z Gita

Next article —

<!--:pl-->Testy - nie takie straszne, jak je malują<!--:--><!--:en-->Tests - not so scary as they used to<!--:-->

You May Also Like
DDevelopment & Design

Apache HISE + Apache Camel

Check out this SlideShare Presentation: Apache HISE + Apache CamelView more presentations from Rafal Rusin.Check out this SlideShare Presentation: Apache HISE + Apache CamelView more presentations from Rafal Rusin.
DDevelopment & Design

Sample for lift-ng: Micro-burn 1.0.0 released

During a last few evenings in my free time I've worked on mini-application called micro-burn. The idea of it appear from work with Agile Jira in our commercial project. This is a great tool for agile projects management. It has inline tasks edition, drag & drop board, reports and many more, but it also have a few drawbacks that turn down our team motivation.

Motivation

From time to time our sprints scope is changing. It is not a big deal because we are trying to be agile :-) but Jira's burndowchart in this situation draw a peek. Because in fact that chart shows scope changes not a real burndown. It means, that chart cannot break down an x-axis if we really do more than we were planned – it always stop on at most zero.

Also for better progress monitoring we've started to split our user stories to technical tasks and estimating them. Original burndowchart doesn't show points from technical tasks. I can find motivation of this – user story almost finished isn't finished at all until user can use it. But in the other hand, if we know which tasks is problematic we can do some teamwork to move it on.

So I realize that it is a good opportunity to try some new approaches and tools.

Tools

I've started with lift framework. In the World of Single Page Applications, this framework has more than simple interface for serving REST services. It comes with awesome Comet support. Comet is a replacement for WebSockets that run on all browsers. It supports long polling and transparent fallback to short polling if limit of client connections exceed. In backend you can handle pushes in CometActor. For further reading take a look at Roundtrip promises

But lift framework is also a kind of framework of frameworks. You can handle own abstraction of CometActors and push to client javascript that shorten up your way from server to client. So it was the trigger for author of lift-ng to make a lift with Angular integration that is build on top of lift. It provides AngularActors from which you can emit/broadcast events to scope of controller. NgModelBinders that synchronize your backend model with client scope in a few lines! I've used them to send project state (all sprints and thier details) to client and notify him about scrum board changes. My actor doing all of this hard work looks pretty small:

Lift-ng also provides factories for creating of Angular services. Services could respond with futures that are transformed to Angular promises in-fly. This is all what was need to serve sprint history:

And on the client side - use of service:


In my opinion this two frameworks gives a huge boost in developing of web applications. You have the power of strongly typing with Scala, you can design your domain on Actors and all of this with simplicity of node.js – lack of json trasforming boilerplate and dynamic application reload.

DDD + Event Sourcing

I've also tried a few fresh approaches to DDD. I've organize domain objects in actors. There are SprintActors with encapsulate sprint aggregate root. Task changes are stored as events which are computed as a difference between two boards states. When it should be provided a history of sprint, next board states are computed from initial state and sequence of events. So I realize that the best way to keep this kind of event sourcing approach tested is to make random tests. This is a test doing random changes at board, calculating events and checking if initial state + events is equals to previously created state:



First look

Screenshot of first version:


If you want to look at this closer, check the source code or download ready to run fatjar on github.During a last few evenings in my free time I've worked on mini-application called micro-burn. The idea of it appear from work with Agile Jira in our commercial project. This is a great tool for agile projects management. It has inline tasks edition, drag & drop board, reports and many more, but it also have a few drawbacks that turn down our team motivation.