You are browsing a read-only backup copy of Wikitech. The live site can be found at wikitech.wikimedia.org
Kerberos is a computer-network authentication protocol that works on the basis of tickets to allow nodes communicating over a non-secure network to prove their identity to one another in a secure manner. The protocol was named after the character Kerberos (or Cerberus) from Greek mythology, the ferocious three-headed guard dog of Hades.
We use it mostly for Hadoop, since it is the only reliable way to add proper authentication for users/daemons/etc..
If you are interested in using Kerberos with Hadoop, please check Analytics/Systems/Kerberos/UserGuide. The rest of the sections are more meant to illustrate how the Kerberos infrastructure is set up.
We have two hosts:
Both hosts are running a kdc daemon, meanwhile only one kadmind and kpropd. In puppet we set up the master and its replica, usually 1001 and 2001 respectively.
Daemons and their roles
- kdc (Key Distribution Centre) - a daemon that authenticate a principal returning a TGT (Ticket Granting Ticket) and acting as TGS (Ticket Granting Service)
- kadmind - a daemon that offers a cli to modify/read the Kerberos database (add principals, generate keytabs, etc..). Runs only on the master kdc node.
- kprop - a cli util that runs periodically on the master node (via systemd timer) propagating changes to the database to the replicas. Runs only on the master kdc node.
- kpropd - a daemon running only on all the replicas, able to get updates sent by kprop and apply the to the local database.
Bootstrap a kerberos host
There are two roles defined in puppet for a kerberos node:
The most common operation should be bootstrapping a new replica, but if you are building a cluster from scratch a new master will need to be created as well.
The master node is the one that holds the authoritative kerberos database containing the credentials for a certain realm. The only realm supported up to now is the WIKIMEDIA one. One you have applied the role::kerberos::kdc to the master node, everything should install/configure nicely but the kerberos daemons shouldn't be able to start correctly due to the absence of a properly bootstrapped database. In order to create a database, execute the following on the master node:
sudo kdb5_util create -s -r WIKIMEDIA
You will have to enter a password. The one of the WIKIMEDIA realm is stored in pwstore.
At this point you should be able to start the kdc and kadmind daemons.
The replica node needs to be bootstrapped as the master node, but it also need something more to allow replication with kprop/kpropd to work. Everything should already be deployed by puppet, but there are some manual steps needed:
- log on the master node (not the replica, it is important) and create host principals for the replica. For example:
- Generate principal and keytab for the replica, and create the keytab in a tmp location
elukey@krb1001:~$ sudo kadmin.local Authenticating as principal host/admin@WIKIMEDIA with password. kadmin.local: addprinc -randkey host/krb2001.codfw.wmnet@WIKIMEDIA kadmin.local: ktadd -k /tmp/krb2001.keytab host/krb2001.codfw.wmnet@WIKIMEDIA
- Stop krb5-kdc and krb5-admin-server on the replica host
- Copy the
/tmp/krb2001.keytab keytabto the replica node, and rename it
- Restart the daemons
At this point it should be possible to execute the
replicate-krb-database.service on the master node to see if database replication works correctly. The service is a very basic script that dumps the current status of the Kerberos database and sends it to each replica listed in hiera via kprop. On Every replica the kpropd daemon should be listening on port 754 to get updates from the master (puppet should take care of setting up kpropd with its related acls).
Good docs to use as reference
There are two main workflows to save the Kerberos database data:
- replicate-krb-database (systemd timer) that periodically dumps the database and replicates to all the replicas via kprop.
- backup-kdc-database (systemd timer), running on all nodes, that periodically takes a snapshot of the local database and save it under /srv.
Why do we need two? If the master database gets corrupted/inconsistent and replicated to all the replicas, then we end up having the wrong data everywhere. Having a local list of database snapshots (especially on the master) should help in having one good version of the database to restore.
The master db snapshots are also saved periodically in Bacula.
Handling failures and failover
In the ideal scenario, when nothing is broken or under maintenance, this is the configuration:
- krb1001 is the master KDC node. It runs the kdc and the kadmind daemons, together with the replicate-krb-database timer. Any credential (principal, keytab) management (add/remove/etc..) should only be done on this host via the kadmin.local CLI tool. A rsync server allows the copy of new keytabs from puppetmaster1001 (in order to add them to the puppet private repo).
- krb2001 is one of the KDC node replicas. It runs the kdc daemon, but not any kadmind, and also the kpropd daemon.
As explained in the above sections, any kerberos client has a list of all the kdc daemons available and can failover transparently if a host is suddenly unreachable. Replication is trickier, since:
- on krb1001, the replicate-krb-database timer periodically dumps the content of the Kerberos database to file and use the kprop (please do not confuse it with kpropd) tool to send the new database snapshot to all the replica hosts.
- on krb2001, one of the replicas, the kpropd listens on port 754 for updates coming from krb1001 and apply them to the local copy of the database if needed.
What happens is krb2001 goes down? Not much, the clients will not notice it, but hiera will need to be changed to avoid replication to fail (namely removing krb2001 from the list of kdc hosts).
What happens if krb1001 goes down? This use case is trickier and needs to be reviewed:
- The hiera configuration needs to be changed to remove krb1001 from the available kdc hosts, and set krb2001 as the active kadmind master.
- In this way, all the configuration available on krb1001 to manage credentials and rsync then to puppetmaster1001 will be added to krb2001.
- The hiera configuration related to the list of kdc servers needs to be updated, since the clients will try to contact the kdc hosts in the order that the config specify. In this case, changing a password on krb2001 (new master) having krb1001 listed as first kdc server will cause failures in logging for the related principal until the new credentials are propagated to krb1001 via replicate-krb-database.
The above step is tricky since we have to remind that every kerberos host has its local database with credentials. The master node uses kprop to replicate credentials, so a client can hit any kdc daemon and obtain a consistent authentication. But if the master node changes, and credentials are modified on the new host, they need to be propagated back when the old master node is ready to get back its leadership status (say after maintenance). This should be handled transparently by puppet when changing the hiera configuration, but some care is needed by the operator to triple check that everything happens correctly.
Manage principals and keytabs
In the Kerberos world, every user is called 'principal' and it may be related to a real human user or a service. In the former case, the user simply kinit, input a password and receive a TGT (that will be used later on to get service access etc..). In the latter, there is a convenient way to store the password to allow a transparent service bootstrap without any manual password input by the operator, namely the keytab.
Create a principal for a real user
In this case, the workflow is to generate the principal with a temporary password, and then allow the user to change it as soon as possible. This is currently possible logging on the current kadmin primary (see Hiera
kerberos_kadmin_server_primar or just log in in one of the krb nodes, there is a bit motd that explains what servers should not be used) and using the command
manage_principals.py like the following:
- sudo manage_principals.py get elukey
- sudo manage_principals.py create elukey --firstname.lastname@example.org
The second use case is very interesting, since it handles the workflow described above:
- A principal is created via kadmin.local.
- A temporary random password is set, together with its expiry in seconds.
- An email is sent to the email_address provided, with instructions about how to change the password. It is sufficient to log in on a node that is a kerberos client, run kinit and follow the instructions.
Create a keytab for a service
In order to properly authenticate a service, we need to:
- create the service principal (example:
- create the keytab for the principal (that eventually will be stored on the the host to be used by a daemon/service).
There is a script on the krb nodes that can help with this:
generate_keytabs.py. It takes as input a file listing, for every row, what to create (principals, keytabs, etc..) and it stores the output (e.g. keytab files) to a known location by default under /srv.
The keytabs are deployed in puppet via the secret module (see profile::kerberos::keytabs) so they need to be copied to puppetmaster1001 in some way. There is a handy rsync server on the krb hosts that offers a module to copy srv only from puppetmaster1001, using user+password (see /srv/kerberos/rsync_secrets_file on the krb hosts). For example, you can do on puppetmaster1001:
mkdir /home/elukey/keytabs rsync email@example.com::srv-keytabs ./keytabs chown -R gitpuppet:gitpuppet ./keytabs sudo cp -a ./keytabs/* /srv/private/modules/secret/secrets/kerberos/keytabs/ rm -rf /home/elukey/keytabs
Another option is of course to rsync directly to the right location in the private repo and chown -R gitpuppet:gitpuppet.
List of principals and their keytabs for Hadoop
The following configuration files are the ones used for the Hadoop test cluster. They are meant to be used with the
Hadoop Master node
analytics1028.eqiad.wmnet,create_princ,HTTP analytics1028.eqiad.wmnet,create_princ,hdfs analytics1028.eqiad.wmnet,create_princ,yarn analytics1028.eqiad.wmnet,create_princ,mapred analytics1028.eqiad.wmnet,create_keytab,HTTP analytics1028.eqiad.wmnet,create_keytab,hdfs analytics1028.eqiad.wmnet,create_keytab,yarn analytics1028.eqiad.wmnet,create_keytab,mapred
Hadoop Master (standby) node
analytics1029.eqiad.wmnet,create_princ,HTTP analytics1029.eqiad.wmnet,create_princ,hdfs analytics1029.eqiad.wmnet,create_princ,yarn analytics1029.eqiad.wmnet,create_princ,mapred analytics1029.eqiad.wmnet,create_keytab,HTTP analytics1029.eqiad.wmnet,create_keytab,hdfs analytics1029.eqiad.wmnet,create_keytab,yarn analytics1029.eqiad.wmnet,create_keytab,mapred
Hadoop worker node
analytics1031.eqiad.wmnet,create_princ,HTTP analytics1031.eqiad.wmnet,create_princ,hdfs analytics1031.eqiad.wmnet,create_princ,yarn analytics1031.eqiad.wmnet,create_keytab,yarn analytics1031.eqiad.wmnet,create_keytab,hdfs analytics1031.eqiad.wmnet,create_keytab,HTTP
Hadoop coordinator (please note the special need for Oozie to have a single keytab for two principals)
analytics1030.eqiad.wmnet,create_princ,HTTP analytics1030.eqiad.wmnet,create_princ,hdfs analytics1030.eqiad.wmnet,create_princ,yarn analytics1030.eqiad.wmnet,create_princ,hive analytics1030.eqiad.wmnet,create_princ,oozie analytics1030.eqiad.wmnet,create_princ,HTTP-oozie analytics1030.eqiad.wmnet,create_princ,analytics analytics1030.eqiad.wmnet,create_keytab,HTTP analytics1030.eqiad.wmnet,create_keytab,hdfs analytics1030.eqiad.wmnet,create_keytab,yarn analytics1030.eqiad.wmnet,create_keytab,hive analytics1030.eqiad.wmnet,create_keytab,oozie analytics1030.eqiad.wmnet,create_keytab,HTTP-oozie analytics1030.eqiad.wmnet,merge_keytab,HTTP-oozie,oozie analytics1030.eqiad.wmnet,create_keytab,analytics
Druid worker (only kerberos credentials to access HDFS, not to authenticate Druid's clients)
analytics1039.eqiad.wmnet,create_princ,HTTP analytics1039.eqiad.wmnet,create_princ,hue analytics1039.eqiad.wmnet,create_keytab,HTTP analytics1039.eqiad.wmnet,create_keytab,hue