Changing the Kerberos principal for Apache Zookeeper causes an
AuthFailedException that must be fixed on the client side with
In July of 2015, while at Avalon Consulting, LLC, I spent two weeks working with a company to secure their Hortonworks Data Platform (HDP) cluster. This included SSL, LDAP, and Kerberos. One of the factors that required some additional effort was the requirement that Linux usernames must be 8 characters, which affected Kerberos principal length as well. This requirement was enforced by Centrify but this could have been enforced through other means. The other day I found that the same problem exists today ~3+ years after I had successfully completed the security implementation. Hopefully this blog helps avoid future headaches when dealing with changing principals, especially Apache Zookeeper.
Effects of Changing the Apache Zookeeper Principal
Apache Ambari during the “Enable Kerberos” wizard does permit changing the principals that get generated. This customer wanted 8 character user names so we changed the corresponding principals to follow that requirement. The Kerberos wizard completed succcessully up until the service start portion. Many of the services started correctly but there were a few that wouldn’t come up.
The few I remember that failed were HDFS ZKFC, Apache HBase, and Apache Storm. The common trend was that it was any service that had to connect to Apache Zookeeper. There were errors in the logs about not being able to authenticate to Zookeeper due to
Example Zookeeper client error message:
org.apache.zookeeper.KeeperException$AuthFailedException: KeeperErrorCode = AuthFailed for ...
Fixing Zookeeper Clients for a Custom Zookeeper Principal
The root cause of the problem after some digging was that Zookeeper assumed that the service principal was
zookeeper. ZOOKEEPER-1811 changed the hardcoded
zookeeper string to a property that could be configured. The new property
zookeeper.sasl.client.username allowed a user to adjust on the client side the principal. Since this a client side setting, all clients of Zookeeper needed to have this property configured. The system property that needed to be set is shown below.
For Apache Zookeeper since it is a client of itself, the
JVMFLAGS property must be set in
Zookeeper Env -> client_opts or on the command line
JVMFLAGS="-Dsun.security.krb5.debug=true -Dzookeeper.sasl.client.username=zk" /usr/hdp/current/zookeeper-client/bin/zkCli.sh -server SERVERNAME
For Apache Hadoop HDFS, the system property must be added to
Hadoop Env -> HADOOP_OPTS. For Apache HBase, the system property must be added to
HBase Env -> HBASE_OPTS. For Apache Storm, the system property must be added to
worker.childopts. At the time, Ambari Infra Solr, Ambari Metrics, and Kafka didn’t exist but the same system property would be needed as described here and here.
Current Status of Apache Ambari and Custom Principals
I haven’t checked how Apache Ambari handles custom principals in the last year or so. I know there have been significant improvements to how Ambari secures Apache Hadoop clusters. Apache Ambari 2.5.0 added Zookeeper ACLs and authentication for a bunch of services. According to this post the problem described is still a problem with HDP 2.6.x and Ambari 2.5.x. Ambari 2.6.x is out but could have the same issues.
Regardless of how improved Apache Ambari becomes, I would suggest not to change the principals for Zookeeper or other services since it could cause a lot of problems.